arc.c revision 286764
1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22219089Spjd * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23277826Sdelphij * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24268123Sdelphij * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25260835Sdelphij * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26286764Smav * Copyright 2015 Nexenta Systems, Inc. All rights reserved. 27168404Spjd */ 28168404Spjd 29168404Spjd/* 30168404Spjd * DVA-based Adjustable Replacement Cache 31168404Spjd * 32168404Spjd * While much of the theory of operation used here is 33168404Spjd * based on the self-tuning, low overhead replacement cache 34168404Spjd * presented by Megiddo and Modha at FAST 2003, there are some 35168404Spjd * significant differences: 36168404Spjd * 37168404Spjd * 1. The Megiddo and Modha model assumes any page is evictable. 38168404Spjd * Pages in its cache cannot be "locked" into memory. This makes 39168404Spjd * the eviction algorithm simple: evict the last page in the list. 40168404Spjd * This also make the performance characteristics easy to reason 41168404Spjd * about. Our cache is not so simple. At any given moment, some 42168404Spjd * subset of the blocks in the cache are un-evictable because we 43168404Spjd * have handed out a reference to them. Blocks are only evictable 44168404Spjd * when there are no external references active. This makes 45168404Spjd * eviction far more problematic: we choose to evict the evictable 46168404Spjd * blocks that are the "lowest" in the list. 47168404Spjd * 48168404Spjd * There are times when it is not possible to evict the requested 49168404Spjd * space. In these circumstances we are unable to adjust the cache 50168404Spjd * size. To prevent the cache growing unbounded at these times we 51185029Spjd * implement a "cache throttle" that slows the flow of new data 52185029Spjd * into the cache until we can make space available. 53168404Spjd * 54168404Spjd * 2. The Megiddo and Modha model assumes a fixed cache size. 55168404Spjd * Pages are evicted when the cache is full and there is a cache 56168404Spjd * miss. Our model has a variable sized cache. It grows with 57185029Spjd * high use, but also tries to react to memory pressure from the 58168404Spjd * operating system: decreasing its size when system memory is 59168404Spjd * tight. 60168404Spjd * 61168404Spjd * 3. The Megiddo and Modha model assumes a fixed page size. All 62251631Sdelphij * elements of the cache are therefore exactly the same size. So 63168404Spjd * when adjusting the cache size following a cache miss, its simply 64168404Spjd * a matter of choosing a single page to evict. In our model, we 65168404Spjd * have variable sized cache blocks (rangeing from 512 bytes to 66251631Sdelphij * 128K bytes). We therefore choose a set of blocks to evict to make 67168404Spjd * space for a cache miss that approximates as closely as possible 68168404Spjd * the space used by the new block. 69168404Spjd * 70168404Spjd * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71168404Spjd * by N. Megiddo & D. Modha, FAST 2003 72168404Spjd */ 73168404Spjd 74168404Spjd/* 75168404Spjd * The locking model: 76168404Spjd * 77168404Spjd * A new reference to a cache buffer can be obtained in two 78168404Spjd * ways: 1) via a hash table lookup using the DVA as a key, 79185029Spjd * or 2) via one of the ARC lists. The arc_read() interface 80168404Spjd * uses method 1, while the internal arc algorithms for 81251631Sdelphij * adjusting the cache use method 2. We therefore provide two 82168404Spjd * types of locks: 1) the hash table lock array, and 2) the 83168404Spjd * arc list locks. 84168404Spjd * 85168404Spjd * Buffers do not have their own mutexs, rather they rely on the 86168404Spjd * hash table mutexs for the bulk of their protection (i.e. most 87168404Spjd * fields in the arc_buf_hdr_t are protected by these mutexs). 88168404Spjd * 89168404Spjd * buf_hash_find() returns the appropriate mutex (held) when it 90168404Spjd * locates the requested buffer in the hash table. It returns 91168404Spjd * NULL for the mutex if the buffer was not in the table. 92168404Spjd * 93168404Spjd * buf_hash_remove() expects the appropriate hash mutex to be 94168404Spjd * already held before it is invoked. 95168404Spjd * 96168404Spjd * Each arc state also has a mutex which is used to protect the 97168404Spjd * buffer list associated with the state. When attempting to 98168404Spjd * obtain a hash table lock while holding an arc list lock you 99168404Spjd * must use: mutex_tryenter() to avoid deadlock. Also note that 100168404Spjd * the active state mutex must be held before the ghost state mutex. 101168404Spjd * 102168404Spjd * Arc buffers may have an associated eviction callback function. 103168404Spjd * This function will be invoked prior to removing the buffer (e.g. 104168404Spjd * in arc_do_user_evicts()). Note however that the data associated 105168404Spjd * with the buffer may be evicted prior to the callback. The callback 106168404Spjd * must be made with *no locks held* (to prevent deadlock). Additionally, 107168404Spjd * the users of callbacks must ensure that their private data is 108268858Sdelphij * protected from simultaneous callbacks from arc_clear_callback() 109168404Spjd * and arc_do_user_evicts(). 110168404Spjd * 111168404Spjd * Note that the majority of the performance stats are manipulated 112168404Spjd * with atomic operations. 113185029Spjd * 114286570Smav * The L2ARC uses the l2ad_mtx on each vdev for the following: 115185029Spjd * 116185029Spjd * - L2ARC buflist creation 117185029Spjd * - L2ARC buflist eviction 118185029Spjd * - L2ARC write completion, which walks L2ARC buflists 119185029Spjd * - ARC header destruction, as it removes from L2ARC buflists 120185029Spjd * - ARC header release, as it removes from L2ARC buflists 121168404Spjd */ 122168404Spjd 123168404Spjd#include <sys/spa.h> 124168404Spjd#include <sys/zio.h> 125251478Sdelphij#include <sys/zio_compress.h> 126168404Spjd#include <sys/zfs_context.h> 127168404Spjd#include <sys/arc.h> 128168404Spjd#include <sys/refcount.h> 129185029Spjd#include <sys/vdev.h> 130219089Spjd#include <sys/vdev_impl.h> 131258632Savg#include <sys/dsl_pool.h> 132286763Smav#include <sys/multilist.h> 133168404Spjd#ifdef _KERNEL 134168404Spjd#include <sys/dnlc.h> 135168404Spjd#endif 136168404Spjd#include <sys/callb.h> 137168404Spjd#include <sys/kstat.h> 138248572Ssmh#include <sys/trim_map.h> 139219089Spjd#include <zfs_fletcher.h> 140168404Spjd#include <sys/sdt.h> 141168404Spjd 142191902Skmacy#include <vm/vm_pageout.h> 143272483Ssmh#include <machine/vmparam.h> 144191902Skmacy 145240133Smm#ifdef illumos 146240133Smm#ifndef _KERNEL 147240133Smm/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 148240133Smmboolean_t arc_watch = B_FALSE; 149240133Smmint arc_procfd; 150240133Smm#endif 151240133Smm#endif /* illumos */ 152240133Smm 153286763Smavstatic kmutex_t arc_reclaim_lock; 154286763Smavstatic kcondvar_t arc_reclaim_thread_cv; 155286763Smavstatic boolean_t arc_reclaim_thread_exit; 156286763Smavstatic kcondvar_t arc_reclaim_waiters_cv; 157168404Spjd 158286763Smavstatic kmutex_t arc_user_evicts_lock; 159286763Smavstatic kcondvar_t arc_user_evicts_cv; 160286763Smavstatic boolean_t arc_user_evicts_thread_exit; 161286763Smav 162286625Smavuint_t arc_reduce_dnlc_percent = 3; 163168404Spjd 164258632Savg/* 165286763Smav * The number of headers to evict in arc_evict_state_impl() before 166286763Smav * dropping the sublist lock and evicting from another sublist. A lower 167286763Smav * value means we're more likely to evict the "correct" header (i.e. the 168286763Smav * oldest header in the arc state), but comes with higher overhead 169286763Smav * (i.e. more invocations of arc_evict_state_impl()). 170258632Savg */ 171286763Smavint zfs_arc_evict_batch_limit = 10; 172258632Savg 173286763Smav/* 174286763Smav * The number of sublists used for each of the arc state lists. If this 175286763Smav * is not set to a suitable value by the user, it will be configured to 176286763Smav * the number of CPUs on the system in arc_init(). 177286763Smav */ 178286763Smavint zfs_arc_num_sublists_per_state = 0; 179286763Smav 180168404Spjd/* number of seconds before growing cache again */ 181168404Spjdstatic int arc_grow_retry = 60; 182168404Spjd 183286763Smav/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 184286763Smavint zfs_arc_overflow_shift = 8; 185286763Smav 186208373Smm/* shift of arc_c for calculating both min and max arc_p */ 187208373Smmstatic int arc_p_min_shift = 4; 188208373Smm 189208373Smm/* log2(fraction of arc to reclaim) */ 190286625Smavstatic int arc_shrink_shift = 7; 191208373Smm 192168404Spjd/* 193286625Smav * log2(fraction of ARC which must be free to allow growing). 194286625Smav * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 195286625Smav * when reading a new block into the ARC, we will evict an equal-sized block 196286625Smav * from the ARC. 197286625Smav * 198286625Smav * This must be less than arc_shrink_shift, so that when we shrink the ARC, 199286625Smav * we will still not allow it to grow. 200286625Smav */ 201286625Smavint arc_no_grow_shift = 5; 202286625Smav 203286625Smav 204286625Smav/* 205168404Spjd * minimum lifespan of a prefetch block in clock ticks 206168404Spjd * (initialized in arc_init()) 207168404Spjd */ 208168404Spjdstatic int arc_min_prefetch_lifespan; 209168404Spjd 210258632Savg/* 211258632Savg * If this percent of memory is free, don't throttle. 212258632Savg */ 213258632Savgint arc_lotsfree_percent = 10; 214258632Savg 215208373Smmstatic int arc_dead; 216194043Skmacyextern int zfs_prefetch_disable; 217168404Spjd 218168404Spjd/* 219185029Spjd * The arc has filled available memory and has now warmed up. 220185029Spjd */ 221185029Spjdstatic boolean_t arc_warm; 222185029Spjd 223286762Smav/* 224286762Smav * These tunables are for performance analysis. 225286762Smav */ 226185029Spjduint64_t zfs_arc_max; 227185029Spjduint64_t zfs_arc_min; 228185029Spjduint64_t zfs_arc_meta_limit = 0; 229275780Sdelphijuint64_t zfs_arc_meta_min = 0; 230208373Smmint zfs_arc_grow_retry = 0; 231208373Smmint zfs_arc_shrink_shift = 0; 232208373Smmint zfs_arc_p_min_shift = 0; 233242845Sdelphijint zfs_disable_dup_eviction = 0; 234269230Sdelphijuint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 235272483Ssmhu_int zfs_arc_free_target = 0; 236185029Spjd 237270759Ssmhstatic int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 238275748Sdelphijstatic int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 239270759Ssmh 240270759Ssmh#ifdef _KERNEL 241270759Ssmhstatic void 242270759Ssmharc_free_target_init(void *unused __unused) 243270759Ssmh{ 244270759Ssmh 245272483Ssmh zfs_arc_free_target = vm_pageout_wakeup_thresh; 246270759Ssmh} 247270759SsmhSYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 248270759Ssmh arc_free_target_init, NULL); 249270759Ssmh 250185029SpjdTUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 251275780SdelphijTUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 252273026SdelphijTUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 253168473SpjdSYSCTL_DECL(_vfs_zfs); 254217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 255168473Spjd "Maximum ARC size"); 256217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 257168473Spjd "Minimum ARC size"); 258269230SdelphijSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 259269230Sdelphij &zfs_arc_average_blocksize, 0, 260269230Sdelphij "ARC average blocksize"); 261273026SdelphijSYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 262273026Sdelphij &arc_shrink_shift, 0, 263273026Sdelphij "log2(fraction of arc to reclaim)"); 264273026Sdelphij 265270759Ssmh/* 266270759Ssmh * We don't have a tunable for arc_free_target due to the dependency on 267270759Ssmh * pagedaemon initialisation. 268270759Ssmh */ 269270759SsmhSYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 270270759Ssmh CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 271270759Ssmh sysctl_vfs_zfs_arc_free_target, "IU", 272270759Ssmh "Desired number of free pages below which ARC triggers reclaim"); 273168404Spjd 274270759Ssmhstatic int 275270759Ssmhsysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 276270759Ssmh{ 277270759Ssmh u_int val; 278270759Ssmh int err; 279270759Ssmh 280270759Ssmh val = zfs_arc_free_target; 281270759Ssmh err = sysctl_handle_int(oidp, &val, 0, req); 282270759Ssmh if (err != 0 || req->newptr == NULL) 283270759Ssmh return (err); 284270759Ssmh 285272483Ssmh if (val < minfree) 286270759Ssmh return (EINVAL); 287272483Ssmh if (val > vm_cnt.v_page_count) 288270759Ssmh return (EINVAL); 289270759Ssmh 290270759Ssmh zfs_arc_free_target = val; 291270759Ssmh 292270759Ssmh return (0); 293270759Ssmh} 294275748Sdelphij 295275748Sdelphij/* 296275748Sdelphij * Must be declared here, before the definition of corresponding kstat 297275748Sdelphij * macro which uses the same names will confuse the compiler. 298275748Sdelphij */ 299275748SdelphijSYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 300275748Sdelphij CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 301275748Sdelphij sysctl_vfs_zfs_arc_meta_limit, "QU", 302275748Sdelphij "ARC metadata limit"); 303272483Ssmh#endif 304270759Ssmh 305168404Spjd/* 306185029Spjd * Note that buffers can be in one of 6 states: 307168404Spjd * ARC_anon - anonymous (discussed below) 308168404Spjd * ARC_mru - recently used, currently cached 309168404Spjd * ARC_mru_ghost - recentely used, no longer in cache 310168404Spjd * ARC_mfu - frequently used, currently cached 311168404Spjd * ARC_mfu_ghost - frequently used, no longer in cache 312185029Spjd * ARC_l2c_only - exists in L2ARC but not other states 313185029Spjd * When there are no active references to the buffer, they are 314185029Spjd * are linked onto a list in one of these arc states. These are 315185029Spjd * the only buffers that can be evicted or deleted. Within each 316185029Spjd * state there are multiple lists, one for meta-data and one for 317185029Spjd * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 318185029Spjd * etc.) is tracked separately so that it can be managed more 319185029Spjd * explicitly: favored over data, limited explicitly. 320168404Spjd * 321168404Spjd * Anonymous buffers are buffers that are not associated with 322168404Spjd * a DVA. These are buffers that hold dirty block copies 323168404Spjd * before they are written to stable storage. By definition, 324168404Spjd * they are "ref'd" and are considered part of arc_mru 325168404Spjd * that cannot be freed. Generally, they will aquire a DVA 326168404Spjd * as they are written and migrate onto the arc_mru list. 327185029Spjd * 328185029Spjd * The ARC_l2c_only state is for buffers that are in the second 329185029Spjd * level ARC but no longer in any of the ARC_m* lists. The second 330185029Spjd * level ARC itself may also contain buffers that are in any of 331185029Spjd * the ARC_m* states - meaning that a buffer can exist in two 332185029Spjd * places. The reason for the ARC_l2c_only state is to keep the 333185029Spjd * buffer header in the hash table, so that reads that hit the 334185029Spjd * second level ARC benefit from these fast lookups. 335168404Spjd */ 336168404Spjd 337168404Spjdtypedef struct arc_state { 338286763Smav /* 339286763Smav * list of evictable buffers 340286763Smav */ 341286763Smav multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 342286763Smav /* 343286763Smav * total amount of evictable data in this state 344286763Smav */ 345286763Smav uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; 346286763Smav /* 347286763Smav * total amount of data in this state; this includes: evictable, 348286763Smav * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 349286763Smav */ 350286763Smav uint64_t arcs_size; 351168404Spjd} arc_state_t; 352168404Spjd 353185029Spjd/* The 6 states: */ 354168404Spjdstatic arc_state_t ARC_anon; 355168404Spjdstatic arc_state_t ARC_mru; 356168404Spjdstatic arc_state_t ARC_mru_ghost; 357168404Spjdstatic arc_state_t ARC_mfu; 358168404Spjdstatic arc_state_t ARC_mfu_ghost; 359185029Spjdstatic arc_state_t ARC_l2c_only; 360168404Spjd 361168404Spjdtypedef struct arc_stats { 362168404Spjd kstat_named_t arcstat_hits; 363168404Spjd kstat_named_t arcstat_misses; 364168404Spjd kstat_named_t arcstat_demand_data_hits; 365168404Spjd kstat_named_t arcstat_demand_data_misses; 366168404Spjd kstat_named_t arcstat_demand_metadata_hits; 367168404Spjd kstat_named_t arcstat_demand_metadata_misses; 368168404Spjd kstat_named_t arcstat_prefetch_data_hits; 369168404Spjd kstat_named_t arcstat_prefetch_data_misses; 370168404Spjd kstat_named_t arcstat_prefetch_metadata_hits; 371168404Spjd kstat_named_t arcstat_prefetch_metadata_misses; 372168404Spjd kstat_named_t arcstat_mru_hits; 373168404Spjd kstat_named_t arcstat_mru_ghost_hits; 374168404Spjd kstat_named_t arcstat_mfu_hits; 375168404Spjd kstat_named_t arcstat_mfu_ghost_hits; 376205231Skmacy kstat_named_t arcstat_allocated; 377168404Spjd kstat_named_t arcstat_deleted; 378251629Sdelphij /* 379251629Sdelphij * Number of buffers that could not be evicted because the hash lock 380251629Sdelphij * was held by another thread. The lock may not necessarily be held 381251629Sdelphij * by something using the same buffer, since hash locks are shared 382251629Sdelphij * by multiple buffers. 383251629Sdelphij */ 384168404Spjd kstat_named_t arcstat_mutex_miss; 385251629Sdelphij /* 386251629Sdelphij * Number of buffers skipped because they have I/O in progress, are 387251629Sdelphij * indrect prefetch buffers that have not lived long enough, or are 388251629Sdelphij * not from the spa we're trying to evict from. 389251629Sdelphij */ 390168404Spjd kstat_named_t arcstat_evict_skip; 391286763Smav /* 392286763Smav * Number of times arc_evict_state() was unable to evict enough 393286763Smav * buffers to reach it's target amount. 394286763Smav */ 395286763Smav kstat_named_t arcstat_evict_not_enough; 396208373Smm kstat_named_t arcstat_evict_l2_cached; 397208373Smm kstat_named_t arcstat_evict_l2_eligible; 398208373Smm kstat_named_t arcstat_evict_l2_ineligible; 399286763Smav kstat_named_t arcstat_evict_l2_skip; 400168404Spjd kstat_named_t arcstat_hash_elements; 401168404Spjd kstat_named_t arcstat_hash_elements_max; 402168404Spjd kstat_named_t arcstat_hash_collisions; 403168404Spjd kstat_named_t arcstat_hash_chains; 404168404Spjd kstat_named_t arcstat_hash_chain_max; 405168404Spjd kstat_named_t arcstat_p; 406168404Spjd kstat_named_t arcstat_c; 407168404Spjd kstat_named_t arcstat_c_min; 408168404Spjd kstat_named_t arcstat_c_max; 409168404Spjd kstat_named_t arcstat_size; 410286574Smav /* 411286574Smav * Number of bytes consumed by internal ARC structures necessary 412286574Smav * for tracking purposes; these structures are not actually 413286574Smav * backed by ARC buffers. This includes arc_buf_hdr_t structures 414286574Smav * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 415286574Smav * caches), and arc_buf_t structures (allocated via arc_buf_t 416286574Smav * cache). 417286574Smav */ 418185029Spjd kstat_named_t arcstat_hdr_size; 419286574Smav /* 420286574Smav * Number of bytes consumed by ARC buffers of type equal to 421286574Smav * ARC_BUFC_DATA. This is generally consumed by buffers backing 422286574Smav * on disk user data (e.g. plain file contents). 423286574Smav */ 424208373Smm kstat_named_t arcstat_data_size; 425286574Smav /* 426286574Smav * Number of bytes consumed by ARC buffers of type equal to 427286574Smav * ARC_BUFC_METADATA. This is generally consumed by buffers 428286574Smav * backing on disk data that is used for internal ZFS 429286574Smav * structures (e.g. ZAP, dnode, indirect blocks, etc). 430286574Smav */ 431286574Smav kstat_named_t arcstat_metadata_size; 432286574Smav /* 433286574Smav * Number of bytes consumed by various buffers and structures 434286574Smav * not actually backed with ARC buffers. This includes bonus 435286574Smav * buffers (allocated directly via zio_buf_* functions), 436286574Smav * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 437286574Smav * cache), and dnode_t structures (allocated via dnode_t cache). 438286574Smav */ 439208373Smm kstat_named_t arcstat_other_size; 440286574Smav /* 441286574Smav * Total number of bytes consumed by ARC buffers residing in the 442286574Smav * arc_anon state. This includes *all* buffers in the arc_anon 443286574Smav * state; e.g. data, metadata, evictable, and unevictable buffers 444286574Smav * are all included in this value. 445286574Smav */ 446286574Smav kstat_named_t arcstat_anon_size; 447286574Smav /* 448286574Smav * Number of bytes consumed by ARC buffers that meet the 449286574Smav * following criteria: backing buffers of type ARC_BUFC_DATA, 450286574Smav * residing in the arc_anon state, and are eligible for eviction 451286574Smav * (e.g. have no outstanding holds on the buffer). 452286574Smav */ 453286574Smav kstat_named_t arcstat_anon_evictable_data; 454286574Smav /* 455286574Smav * Number of bytes consumed by ARC buffers that meet the 456286574Smav * following criteria: backing buffers of type ARC_BUFC_METADATA, 457286574Smav * residing in the arc_anon state, and are eligible for eviction 458286574Smav * (e.g. have no outstanding holds on the buffer). 459286574Smav */ 460286574Smav kstat_named_t arcstat_anon_evictable_metadata; 461286574Smav /* 462286574Smav * Total number of bytes consumed by ARC buffers residing in the 463286574Smav * arc_mru state. This includes *all* buffers in the arc_mru 464286574Smav * state; e.g. data, metadata, evictable, and unevictable buffers 465286574Smav * are all included in this value. 466286574Smav */ 467286574Smav kstat_named_t arcstat_mru_size; 468286574Smav /* 469286574Smav * Number of bytes consumed by ARC buffers that meet the 470286574Smav * following criteria: backing buffers of type ARC_BUFC_DATA, 471286574Smav * residing in the arc_mru state, and are eligible for eviction 472286574Smav * (e.g. have no outstanding holds on the buffer). 473286574Smav */ 474286574Smav kstat_named_t arcstat_mru_evictable_data; 475286574Smav /* 476286574Smav * Number of bytes consumed by ARC buffers that meet the 477286574Smav * following criteria: backing buffers of type ARC_BUFC_METADATA, 478286574Smav * residing in the arc_mru state, and are eligible for eviction 479286574Smav * (e.g. have no outstanding holds on the buffer). 480286574Smav */ 481286574Smav kstat_named_t arcstat_mru_evictable_metadata; 482286574Smav /* 483286574Smav * Total number of bytes that *would have been* consumed by ARC 484286574Smav * buffers in the arc_mru_ghost state. The key thing to note 485286574Smav * here, is the fact that this size doesn't actually indicate 486286574Smav * RAM consumption. The ghost lists only consist of headers and 487286574Smav * don't actually have ARC buffers linked off of these headers. 488286574Smav * Thus, *if* the headers had associated ARC buffers, these 489286574Smav * buffers *would have* consumed this number of bytes. 490286574Smav */ 491286574Smav kstat_named_t arcstat_mru_ghost_size; 492286574Smav /* 493286574Smav * Number of bytes that *would have been* consumed by ARC 494286574Smav * buffers that are eligible for eviction, of type 495286574Smav * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 496286574Smav */ 497286574Smav kstat_named_t arcstat_mru_ghost_evictable_data; 498286574Smav /* 499286574Smav * Number of bytes that *would have been* consumed by ARC 500286574Smav * buffers that are eligible for eviction, of type 501286574Smav * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 502286574Smav */ 503286574Smav kstat_named_t arcstat_mru_ghost_evictable_metadata; 504286574Smav /* 505286574Smav * Total number of bytes consumed by ARC buffers residing in the 506286574Smav * arc_mfu state. This includes *all* buffers in the arc_mfu 507286574Smav * state; e.g. data, metadata, evictable, and unevictable buffers 508286574Smav * are all included in this value. 509286574Smav */ 510286574Smav kstat_named_t arcstat_mfu_size; 511286574Smav /* 512286574Smav * Number of bytes consumed by ARC buffers that are eligible for 513286574Smav * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 514286574Smav * state. 515286574Smav */ 516286574Smav kstat_named_t arcstat_mfu_evictable_data; 517286574Smav /* 518286574Smav * Number of bytes consumed by ARC buffers that are eligible for 519286574Smav * eviction, of type ARC_BUFC_METADATA, and reside in the 520286574Smav * arc_mfu state. 521286574Smav */ 522286574Smav kstat_named_t arcstat_mfu_evictable_metadata; 523286574Smav /* 524286574Smav * Total number of bytes that *would have been* consumed by ARC 525286574Smav * buffers in the arc_mfu_ghost state. See the comment above 526286574Smav * arcstat_mru_ghost_size for more details. 527286574Smav */ 528286574Smav kstat_named_t arcstat_mfu_ghost_size; 529286574Smav /* 530286574Smav * Number of bytes that *would have been* consumed by ARC 531286574Smav * buffers that are eligible for eviction, of type 532286574Smav * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 533286574Smav */ 534286574Smav kstat_named_t arcstat_mfu_ghost_evictable_data; 535286574Smav /* 536286574Smav * Number of bytes that *would have been* consumed by ARC 537286574Smav * buffers that are eligible for eviction, of type 538286574Smav * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 539286574Smav */ 540286574Smav kstat_named_t arcstat_mfu_ghost_evictable_metadata; 541185029Spjd kstat_named_t arcstat_l2_hits; 542185029Spjd kstat_named_t arcstat_l2_misses; 543185029Spjd kstat_named_t arcstat_l2_feeds; 544185029Spjd kstat_named_t arcstat_l2_rw_clash; 545208373Smm kstat_named_t arcstat_l2_read_bytes; 546208373Smm kstat_named_t arcstat_l2_write_bytes; 547185029Spjd kstat_named_t arcstat_l2_writes_sent; 548185029Spjd kstat_named_t arcstat_l2_writes_done; 549185029Spjd kstat_named_t arcstat_l2_writes_error; 550286763Smav kstat_named_t arcstat_l2_writes_lock_retry; 551185029Spjd kstat_named_t arcstat_l2_evict_lock_retry; 552185029Spjd kstat_named_t arcstat_l2_evict_reading; 553286570Smav kstat_named_t arcstat_l2_evict_l1cached; 554185029Spjd kstat_named_t arcstat_l2_free_on_write; 555274172Savg kstat_named_t arcstat_l2_cdata_free_on_write; 556185029Spjd kstat_named_t arcstat_l2_abort_lowmem; 557185029Spjd kstat_named_t arcstat_l2_cksum_bad; 558185029Spjd kstat_named_t arcstat_l2_io_error; 559185029Spjd kstat_named_t arcstat_l2_size; 560251478Sdelphij kstat_named_t arcstat_l2_asize; 561185029Spjd kstat_named_t arcstat_l2_hdr_size; 562251478Sdelphij kstat_named_t arcstat_l2_compress_successes; 563251478Sdelphij kstat_named_t arcstat_l2_compress_zeros; 564251478Sdelphij kstat_named_t arcstat_l2_compress_failures; 565205231Skmacy kstat_named_t arcstat_l2_write_trylock_fail; 566205231Skmacy kstat_named_t arcstat_l2_write_passed_headroom; 567205231Skmacy kstat_named_t arcstat_l2_write_spa_mismatch; 568206796Spjd kstat_named_t arcstat_l2_write_in_l2; 569205231Skmacy kstat_named_t arcstat_l2_write_hdr_io_in_progress; 570205231Skmacy kstat_named_t arcstat_l2_write_not_cacheable; 571205231Skmacy kstat_named_t arcstat_l2_write_full; 572205231Skmacy kstat_named_t arcstat_l2_write_buffer_iter; 573205231Skmacy kstat_named_t arcstat_l2_write_pios; 574205231Skmacy kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 575205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_iter; 576205231Skmacy kstat_named_t arcstat_l2_write_buffer_list_null_iter; 577242845Sdelphij kstat_named_t arcstat_memory_throttle_count; 578242845Sdelphij kstat_named_t arcstat_duplicate_buffers; 579242845Sdelphij kstat_named_t arcstat_duplicate_buffers_size; 580242845Sdelphij kstat_named_t arcstat_duplicate_reads; 581275748Sdelphij kstat_named_t arcstat_meta_used; 582275748Sdelphij kstat_named_t arcstat_meta_limit; 583275748Sdelphij kstat_named_t arcstat_meta_max; 584275780Sdelphij kstat_named_t arcstat_meta_min; 585168404Spjd} arc_stats_t; 586168404Spjd 587168404Spjdstatic arc_stats_t arc_stats = { 588168404Spjd { "hits", KSTAT_DATA_UINT64 }, 589168404Spjd { "misses", KSTAT_DATA_UINT64 }, 590168404Spjd { "demand_data_hits", KSTAT_DATA_UINT64 }, 591168404Spjd { "demand_data_misses", KSTAT_DATA_UINT64 }, 592168404Spjd { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 593168404Spjd { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 594168404Spjd { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 595168404Spjd { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 596168404Spjd { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 597168404Spjd { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 598168404Spjd { "mru_hits", KSTAT_DATA_UINT64 }, 599168404Spjd { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 600168404Spjd { "mfu_hits", KSTAT_DATA_UINT64 }, 601168404Spjd { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 602205231Skmacy { "allocated", KSTAT_DATA_UINT64 }, 603168404Spjd { "deleted", KSTAT_DATA_UINT64 }, 604168404Spjd { "mutex_miss", KSTAT_DATA_UINT64 }, 605168404Spjd { "evict_skip", KSTAT_DATA_UINT64 }, 606286763Smav { "evict_not_enough", KSTAT_DATA_UINT64 }, 607208373Smm { "evict_l2_cached", KSTAT_DATA_UINT64 }, 608208373Smm { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 609208373Smm { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 610286763Smav { "evict_l2_skip", KSTAT_DATA_UINT64 }, 611168404Spjd { "hash_elements", KSTAT_DATA_UINT64 }, 612168404Spjd { "hash_elements_max", KSTAT_DATA_UINT64 }, 613168404Spjd { "hash_collisions", KSTAT_DATA_UINT64 }, 614168404Spjd { "hash_chains", KSTAT_DATA_UINT64 }, 615168404Spjd { "hash_chain_max", KSTAT_DATA_UINT64 }, 616168404Spjd { "p", KSTAT_DATA_UINT64 }, 617168404Spjd { "c", KSTAT_DATA_UINT64 }, 618168404Spjd { "c_min", KSTAT_DATA_UINT64 }, 619168404Spjd { "c_max", KSTAT_DATA_UINT64 }, 620185029Spjd { "size", KSTAT_DATA_UINT64 }, 621185029Spjd { "hdr_size", KSTAT_DATA_UINT64 }, 622208373Smm { "data_size", KSTAT_DATA_UINT64 }, 623286574Smav { "metadata_size", KSTAT_DATA_UINT64 }, 624208373Smm { "other_size", KSTAT_DATA_UINT64 }, 625286574Smav { "anon_size", KSTAT_DATA_UINT64 }, 626286574Smav { "anon_evictable_data", KSTAT_DATA_UINT64 }, 627286574Smav { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 628286574Smav { "mru_size", KSTAT_DATA_UINT64 }, 629286574Smav { "mru_evictable_data", KSTAT_DATA_UINT64 }, 630286574Smav { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 631286574Smav { "mru_ghost_size", KSTAT_DATA_UINT64 }, 632286574Smav { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 633286574Smav { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 634286574Smav { "mfu_size", KSTAT_DATA_UINT64 }, 635286574Smav { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 636286574Smav { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 637286574Smav { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 638286574Smav { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 639286574Smav { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 640185029Spjd { "l2_hits", KSTAT_DATA_UINT64 }, 641185029Spjd { "l2_misses", KSTAT_DATA_UINT64 }, 642185029Spjd { "l2_feeds", KSTAT_DATA_UINT64 }, 643185029Spjd { "l2_rw_clash", KSTAT_DATA_UINT64 }, 644208373Smm { "l2_read_bytes", KSTAT_DATA_UINT64 }, 645208373Smm { "l2_write_bytes", KSTAT_DATA_UINT64 }, 646185029Spjd { "l2_writes_sent", KSTAT_DATA_UINT64 }, 647185029Spjd { "l2_writes_done", KSTAT_DATA_UINT64 }, 648185029Spjd { "l2_writes_error", KSTAT_DATA_UINT64 }, 649286763Smav { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 650185029Spjd { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 651185029Spjd { "l2_evict_reading", KSTAT_DATA_UINT64 }, 652286570Smav { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 653185029Spjd { "l2_free_on_write", KSTAT_DATA_UINT64 }, 654274172Savg { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 655185029Spjd { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 656185029Spjd { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 657185029Spjd { "l2_io_error", KSTAT_DATA_UINT64 }, 658185029Spjd { "l2_size", KSTAT_DATA_UINT64 }, 659251478Sdelphij { "l2_asize", KSTAT_DATA_UINT64 }, 660185029Spjd { "l2_hdr_size", KSTAT_DATA_UINT64 }, 661251478Sdelphij { "l2_compress_successes", KSTAT_DATA_UINT64 }, 662251478Sdelphij { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 663251478Sdelphij { "l2_compress_failures", KSTAT_DATA_UINT64 }, 664206796Spjd { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 665206796Spjd { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 666206796Spjd { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 667206796Spjd { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 668206796Spjd { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 669206796Spjd { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 670206796Spjd { "l2_write_full", KSTAT_DATA_UINT64 }, 671206796Spjd { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 672206796Spjd { "l2_write_pios", KSTAT_DATA_UINT64 }, 673206796Spjd { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 674206796Spjd { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 675242845Sdelphij { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 676242845Sdelphij { "memory_throttle_count", KSTAT_DATA_UINT64 }, 677242845Sdelphij { "duplicate_buffers", KSTAT_DATA_UINT64 }, 678242845Sdelphij { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 679275748Sdelphij { "duplicate_reads", KSTAT_DATA_UINT64 }, 680275748Sdelphij { "arc_meta_used", KSTAT_DATA_UINT64 }, 681275748Sdelphij { "arc_meta_limit", KSTAT_DATA_UINT64 }, 682275780Sdelphij { "arc_meta_max", KSTAT_DATA_UINT64 }, 683275780Sdelphij { "arc_meta_min", KSTAT_DATA_UINT64 } 684168404Spjd}; 685168404Spjd 686168404Spjd#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 687168404Spjd 688168404Spjd#define ARCSTAT_INCR(stat, val) \ 689251631Sdelphij atomic_add_64(&arc_stats.stat.value.ui64, (val)) 690168404Spjd 691206796Spjd#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 692168404Spjd#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 693168404Spjd 694168404Spjd#define ARCSTAT_MAX(stat, val) { \ 695168404Spjd uint64_t m; \ 696168404Spjd while ((val) > (m = arc_stats.stat.value.ui64) && \ 697168404Spjd (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 698168404Spjd continue; \ 699168404Spjd} 700168404Spjd 701168404Spjd#define ARCSTAT_MAXSTAT(stat) \ 702168404Spjd ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 703168404Spjd 704168404Spjd/* 705168404Spjd * We define a macro to allow ARC hits/misses to be easily broken down by 706168404Spjd * two separate conditions, giving a total of four different subtypes for 707168404Spjd * each of hits and misses (so eight statistics total). 708168404Spjd */ 709168404Spjd#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 710168404Spjd if (cond1) { \ 711168404Spjd if (cond2) { \ 712168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 713168404Spjd } else { \ 714168404Spjd ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 715168404Spjd } \ 716168404Spjd } else { \ 717168404Spjd if (cond2) { \ 718168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 719168404Spjd } else { \ 720168404Spjd ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 721168404Spjd } \ 722168404Spjd } 723168404Spjd 724168404Spjdkstat_t *arc_ksp; 725206796Spjdstatic arc_state_t *arc_anon; 726168404Spjdstatic arc_state_t *arc_mru; 727168404Spjdstatic arc_state_t *arc_mru_ghost; 728168404Spjdstatic arc_state_t *arc_mfu; 729168404Spjdstatic arc_state_t *arc_mfu_ghost; 730185029Spjdstatic arc_state_t *arc_l2c_only; 731168404Spjd 732168404Spjd/* 733168404Spjd * There are several ARC variables that are critical to export as kstats -- 734168404Spjd * but we don't want to have to grovel around in the kstat whenever we wish to 735168404Spjd * manipulate them. For these variables, we therefore define them to be in 736168404Spjd * terms of the statistic variable. This assures that we are not introducing 737168404Spjd * the possibility of inconsistency by having shadow copies of the variables, 738168404Spjd * while still allowing the code to be readable. 739168404Spjd */ 740168404Spjd#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 741168404Spjd#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 742168404Spjd#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 743168404Spjd#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 744168404Spjd#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 745275748Sdelphij#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 746275780Sdelphij#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 747275748Sdelphij#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 748275748Sdelphij#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 749168404Spjd 750251478Sdelphij#define L2ARC_IS_VALID_COMPRESS(_c_) \ 751251478Sdelphij ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 752251478Sdelphij 753168404Spjdstatic int arc_no_grow; /* Don't try to grow cache size */ 754168404Spjdstatic uint64_t arc_tempreserve; 755209962Smmstatic uint64_t arc_loaned_bytes; 756168404Spjd 757168404Spjdtypedef struct arc_callback arc_callback_t; 758168404Spjd 759168404Spjdstruct arc_callback { 760168404Spjd void *acb_private; 761168404Spjd arc_done_func_t *acb_done; 762168404Spjd arc_buf_t *acb_buf; 763168404Spjd zio_t *acb_zio_dummy; 764168404Spjd arc_callback_t *acb_next; 765168404Spjd}; 766168404Spjd 767168404Spjdtypedef struct arc_write_callback arc_write_callback_t; 768168404Spjd 769168404Spjdstruct arc_write_callback { 770168404Spjd void *awcb_private; 771168404Spjd arc_done_func_t *awcb_ready; 772258632Savg arc_done_func_t *awcb_physdone; 773168404Spjd arc_done_func_t *awcb_done; 774168404Spjd arc_buf_t *awcb_buf; 775168404Spjd}; 776168404Spjd 777286570Smav/* 778286570Smav * ARC buffers are separated into multiple structs as a memory saving measure: 779286570Smav * - Common fields struct, always defined, and embedded within it: 780286570Smav * - L2-only fields, always allocated but undefined when not in L2ARC 781286570Smav * - L1-only fields, only allocated when in L1ARC 782286570Smav * 783286570Smav * Buffer in L1 Buffer only in L2 784286570Smav * +------------------------+ +------------------------+ 785286570Smav * | arc_buf_hdr_t | | arc_buf_hdr_t | 786286570Smav * | | | | 787286570Smav * | | | | 788286570Smav * | | | | 789286570Smav * +------------------------+ +------------------------+ 790286570Smav * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 791286570Smav * | (undefined if L1-only) | | | 792286570Smav * +------------------------+ +------------------------+ 793286570Smav * | l1arc_buf_hdr_t | 794286570Smav * | | 795286570Smav * | | 796286570Smav * | | 797286570Smav * | | 798286570Smav * +------------------------+ 799286570Smav * 800286570Smav * Because it's possible for the L2ARC to become extremely large, we can wind 801286570Smav * up eating a lot of memory in L2ARC buffer headers, so the size of a header 802286570Smav * is minimized by only allocating the fields necessary for an L1-cached buffer 803286570Smav * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 804286570Smav * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 805286570Smav * words in pointers. arc_hdr_realloc() is used to switch a header between 806286570Smav * these two allocation states. 807286570Smav */ 808286570Smavtypedef struct l1arc_buf_hdr { 809168404Spjd kmutex_t b_freeze_lock; 810286570Smav#ifdef ZFS_DEBUG 811286570Smav /* 812286570Smav * used for debugging wtih kmem_flags - by allocating and freeing 813286570Smav * b_thawed when the buffer is thawed, we get a record of the stack 814286570Smav * trace that thawed it. 815286570Smav */ 816219089Spjd void *b_thawed; 817286570Smav#endif 818168404Spjd 819168404Spjd arc_buf_t *b_buf; 820168404Spjd uint32_t b_datacnt; 821286570Smav /* for waiting on writes to complete */ 822168404Spjd kcondvar_t b_cv; 823168404Spjd 824168404Spjd /* protected by arc state mutex */ 825168404Spjd arc_state_t *b_state; 826286763Smav multilist_node_t b_arc_node; 827168404Spjd 828168404Spjd /* updated atomically */ 829168404Spjd clock_t b_arc_access; 830168404Spjd 831168404Spjd /* self protecting */ 832168404Spjd refcount_t b_refcnt; 833185029Spjd 834286570Smav arc_callback_t *b_acb; 835286570Smav /* temporary buffer holder for in-flight compressed data */ 836286570Smav void *b_tmp_cdata; 837286570Smav} l1arc_buf_hdr_t; 838286570Smav 839286570Smavtypedef struct l2arc_dev l2arc_dev_t; 840286570Smav 841286570Smavtypedef struct l2arc_buf_hdr { 842286570Smav /* protected by arc_buf_hdr mutex */ 843286570Smav l2arc_dev_t *b_dev; /* L2ARC device */ 844286570Smav uint64_t b_daddr; /* disk address, offset byte */ 845286570Smav /* real alloc'd buffer size depending on b_compress applied */ 846286570Smav int32_t b_asize; 847286570Smav 848185029Spjd list_node_t b_l2node; 849286570Smav} l2arc_buf_hdr_t; 850286570Smav 851286570Smavstruct arc_buf_hdr { 852286570Smav /* protected by hash lock */ 853286570Smav dva_t b_dva; 854286570Smav uint64_t b_birth; 855286570Smav /* 856286570Smav * Even though this checksum is only set/verified when a buffer is in 857286570Smav * the L1 cache, it needs to be in the set of common fields because it 858286570Smav * must be preserved from the time before a buffer is written out to 859286570Smav * L2ARC until after it is read back in. 860286570Smav */ 861286570Smav zio_cksum_t *b_freeze_cksum; 862286570Smav 863286570Smav arc_buf_hdr_t *b_hash_next; 864286570Smav arc_flags_t b_flags; 865286570Smav 866286570Smav /* immutable */ 867286570Smav int32_t b_size; 868286570Smav uint64_t b_spa; 869286570Smav 870286570Smav /* L2ARC fields. Undefined when not in L2ARC. */ 871286570Smav l2arc_buf_hdr_t b_l2hdr; 872286570Smav /* L1ARC fields. Undefined when in l2arc_only state */ 873286570Smav l1arc_buf_hdr_t b_l1hdr; 874168404Spjd}; 875168404Spjd 876275748Sdelphij#ifdef _KERNEL 877275748Sdelphijstatic int 878275748Sdelphijsysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 879275748Sdelphij{ 880275748Sdelphij uint64_t val; 881275748Sdelphij int err; 882275748Sdelphij 883275748Sdelphij val = arc_meta_limit; 884275748Sdelphij err = sysctl_handle_64(oidp, &val, 0, req); 885275748Sdelphij if (err != 0 || req->newptr == NULL) 886275748Sdelphij return (err); 887275748Sdelphij 888275748Sdelphij if (val <= 0 || val > arc_c_max) 889275748Sdelphij return (EINVAL); 890275748Sdelphij 891275748Sdelphij arc_meta_limit = val; 892275748Sdelphij return (0); 893275748Sdelphij} 894275748Sdelphij#endif 895275748Sdelphij 896168404Spjdstatic arc_buf_t *arc_eviction_list; 897168404Spjdstatic arc_buf_hdr_t arc_eviction_hdr; 898168404Spjd 899168404Spjd#define GHOST_STATE(state) \ 900185029Spjd ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 901185029Spjd (state) == arc_l2c_only) 902168404Spjd 903275811Sdelphij#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 904275811Sdelphij#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 905275811Sdelphij#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 906275811Sdelphij#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 907275811Sdelphij#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 908275811Sdelphij#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 909286570Smav 910275811Sdelphij#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 911286570Smav#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 912275811Sdelphij#define HDR_L2_READING(hdr) \ 913286570Smav (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 914286570Smav ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 915275811Sdelphij#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 916275811Sdelphij#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 917275811Sdelphij#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 918168404Spjd 919286570Smav#define HDR_ISTYPE_METADATA(hdr) \ 920286570Smav ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 921286570Smav#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 922286570Smav 923286570Smav#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 924286570Smav#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 925286570Smav 926286570Smav/* For storing compression mode in b_flags */ 927286570Smav#define HDR_COMPRESS_OFFSET 24 928286570Smav#define HDR_COMPRESS_NBITS 7 929286570Smav 930286570Smav#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 931286570Smav HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 932286570Smav#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 933286570Smav HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 934286570Smav 935168404Spjd/* 936185029Spjd * Other sizes 937185029Spjd */ 938185029Spjd 939286570Smav#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 940286570Smav#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 941185029Spjd 942185029Spjd/* 943168404Spjd * Hash table routines 944168404Spjd */ 945168404Spjd 946205253Skmacy#define HT_LOCK_PAD CACHE_LINE_SIZE 947168404Spjd 948168404Spjdstruct ht_lock { 949168404Spjd kmutex_t ht_lock; 950168404Spjd#ifdef _KERNEL 951168404Spjd unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 952168404Spjd#endif 953168404Spjd}; 954168404Spjd 955168404Spjd#define BUF_LOCKS 256 956168404Spjdtypedef struct buf_hash_table { 957168404Spjd uint64_t ht_mask; 958168404Spjd arc_buf_hdr_t **ht_table; 959205264Skmacy struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 960168404Spjd} buf_hash_table_t; 961168404Spjd 962168404Spjdstatic buf_hash_table_t buf_hash_table; 963168404Spjd 964168404Spjd#define BUF_HASH_INDEX(spa, dva, birth) \ 965168404Spjd (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 966168404Spjd#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 967168404Spjd#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 968219089Spjd#define HDR_LOCK(hdr) \ 969219089Spjd (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 970168404Spjd 971168404Spjduint64_t zfs_crc64_table[256]; 972168404Spjd 973185029Spjd/* 974185029Spjd * Level 2 ARC 975185029Spjd */ 976185029Spjd 977272707Savg#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 978251478Sdelphij#define L2ARC_HEADROOM 2 /* num of writes */ 979251478Sdelphij/* 980251478Sdelphij * If we discover during ARC scan any buffers to be compressed, we boost 981251478Sdelphij * our headroom for the next scanning cycle by this percentage multiple. 982251478Sdelphij */ 983251478Sdelphij#define L2ARC_HEADROOM_BOOST 200 984208373Smm#define L2ARC_FEED_SECS 1 /* caching interval secs */ 985208373Smm#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 986185029Spjd 987286598Smav/* 988286598Smav * Used to distinguish headers that are being process by 989286598Smav * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 990286598Smav * address. This can happen when the header is added to the l2arc's list 991286598Smav * of buffers to write in the first stage of l2arc_write_buffers(), but 992286598Smav * has not yet been written out which happens in the second stage of 993286598Smav * l2arc_write_buffers(). 994286598Smav */ 995286598Smav#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 996286598Smav 997185029Spjd#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 998185029Spjd#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 999185029Spjd 1000251631Sdelphij/* L2ARC Performance Tunables */ 1001185029Spjduint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1002185029Spjduint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1003185029Spjduint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1004251478Sdelphijuint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1005185029Spjduint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1006208373Smmuint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1007219089Spjdboolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1008208373Smmboolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1009208373Smmboolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1010185029Spjd 1011217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1012205231Skmacy &l2arc_write_max, 0, "max write size"); 1013217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1014205231Skmacy &l2arc_write_boost, 0, "extra write during warmup"); 1015217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1016205231Skmacy &l2arc_headroom, 0, "number of dev writes"); 1017217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1018205231Skmacy &l2arc_feed_secs, 0, "interval seconds"); 1019217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1020208373Smm &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1021205231Skmacy 1022205231SkmacySYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1023205231Skmacy &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1024208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1025208373Smm &l2arc_feed_again, 0, "turbo warmup"); 1026208373SmmSYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1027208373Smm &l2arc_norw, 0, "no reads during writes"); 1028205231Skmacy 1029217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1030205231Skmacy &ARC_anon.arcs_size, 0, "size of anonymous state"); 1031217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1032205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1033217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1034205231Skmacy &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1035205231Skmacy 1036217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1037205231Skmacy &ARC_mru.arcs_size, 0, "size of mru state"); 1038217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1039205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1040217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1041205231Skmacy &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1042205231Skmacy 1043217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1044205231Skmacy &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 1045217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1046205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1047205231Skmacy "size of metadata in mru ghost state"); 1048217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1049205231Skmacy &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1050205231Skmacy "size of data in mru ghost state"); 1051205231Skmacy 1052217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1053205231Skmacy &ARC_mfu.arcs_size, 0, "size of mfu state"); 1054217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1055205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1056217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1057205231Skmacy &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1058205231Skmacy 1059217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1060205231Skmacy &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 1061217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1062205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1063205231Skmacy "size of metadata in mfu ghost state"); 1064217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1065205231Skmacy &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1066205231Skmacy "size of data in mfu ghost state"); 1067205231Skmacy 1068217367SmdfSYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1069205231Skmacy &ARC_l2c_only.arcs_size, 0, "size of mru state"); 1070205231Skmacy 1071185029Spjd/* 1072185029Spjd * L2ARC Internals 1073185029Spjd */ 1074286570Smavstruct l2arc_dev { 1075185029Spjd vdev_t *l2ad_vdev; /* vdev */ 1076185029Spjd spa_t *l2ad_spa; /* spa */ 1077185029Spjd uint64_t l2ad_hand; /* next write location */ 1078185029Spjd uint64_t l2ad_start; /* first addr on device */ 1079185029Spjd uint64_t l2ad_end; /* last addr on device */ 1080185029Spjd boolean_t l2ad_first; /* first sweep through */ 1081208373Smm boolean_t l2ad_writing; /* currently writing */ 1082286570Smav kmutex_t l2ad_mtx; /* lock for buffer list */ 1083286570Smav list_t l2ad_buflist; /* buffer list */ 1084185029Spjd list_node_t l2ad_node; /* device list node */ 1085286598Smav refcount_t l2ad_alloc; /* allocated bytes */ 1086286570Smav}; 1087185029Spjd 1088185029Spjdstatic list_t L2ARC_dev_list; /* device list */ 1089185029Spjdstatic list_t *l2arc_dev_list; /* device list pointer */ 1090185029Spjdstatic kmutex_t l2arc_dev_mtx; /* device list mutex */ 1091185029Spjdstatic l2arc_dev_t *l2arc_dev_last; /* last device used */ 1092185029Spjdstatic list_t L2ARC_free_on_write; /* free after write buf list */ 1093185029Spjdstatic list_t *l2arc_free_on_write; /* free after write list ptr */ 1094185029Spjdstatic kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1095185029Spjdstatic uint64_t l2arc_ndev; /* number of devices */ 1096185029Spjd 1097185029Spjdtypedef struct l2arc_read_callback { 1098251478Sdelphij arc_buf_t *l2rcb_buf; /* read buffer */ 1099251478Sdelphij spa_t *l2rcb_spa; /* spa */ 1100251478Sdelphij blkptr_t l2rcb_bp; /* original blkptr */ 1101268123Sdelphij zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1102251478Sdelphij int l2rcb_flags; /* original flags */ 1103251478Sdelphij enum zio_compress l2rcb_compress; /* applied compress */ 1104185029Spjd} l2arc_read_callback_t; 1105185029Spjd 1106185029Spjdtypedef struct l2arc_write_callback { 1107185029Spjd l2arc_dev_t *l2wcb_dev; /* device info */ 1108185029Spjd arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1109185029Spjd} l2arc_write_callback_t; 1110185029Spjd 1111185029Spjdtypedef struct l2arc_data_free { 1112185029Spjd /* protected by l2arc_free_on_write_mtx */ 1113185029Spjd void *l2df_data; 1114185029Spjd size_t l2df_size; 1115185029Spjd void (*l2df_func)(void *, size_t); 1116185029Spjd list_node_t l2df_list_node; 1117185029Spjd} l2arc_data_free_t; 1118185029Spjd 1119185029Spjdstatic kmutex_t l2arc_feed_thr_lock; 1120185029Spjdstatic kcondvar_t l2arc_feed_thr_cv; 1121185029Spjdstatic uint8_t l2arc_thread_exit; 1122185029Spjd 1123275811Sdelphijstatic void arc_get_data_buf(arc_buf_t *); 1124275811Sdelphijstatic void arc_access(arc_buf_hdr_t *, kmutex_t *); 1125286763Smavstatic boolean_t arc_is_overflowing(); 1126275811Sdelphijstatic void arc_buf_watch(arc_buf_t *); 1127275811Sdelphij 1128286570Smavstatic arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1129286570Smavstatic uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1130286570Smav 1131275811Sdelphijstatic boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1132275811Sdelphijstatic void l2arc_read_done(zio_t *); 1133185029Spjd 1134286570Smavstatic boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1135275811Sdelphijstatic void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1136275811Sdelphijstatic void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1137251478Sdelphij 1138168404Spjdstatic uint64_t 1139209962Smmbuf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1140168404Spjd{ 1141168404Spjd uint8_t *vdva = (uint8_t *)dva; 1142168404Spjd uint64_t crc = -1ULL; 1143168404Spjd int i; 1144168404Spjd 1145168404Spjd ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1146168404Spjd 1147168404Spjd for (i = 0; i < sizeof (dva_t); i++) 1148168404Spjd crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1149168404Spjd 1150209962Smm crc ^= (spa>>8) ^ birth; 1151168404Spjd 1152168404Spjd return (crc); 1153168404Spjd} 1154168404Spjd 1155168404Spjd#define BUF_EMPTY(buf) \ 1156168404Spjd ((buf)->b_dva.dva_word[0] == 0 && \ 1157286570Smav (buf)->b_dva.dva_word[1] == 0) 1158168404Spjd 1159168404Spjd#define BUF_EQUAL(spa, dva, birth, buf) \ 1160168404Spjd ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1161168404Spjd ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1162168404Spjd ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1163168404Spjd 1164219089Spjdstatic void 1165219089Spjdbuf_discard_identity(arc_buf_hdr_t *hdr) 1166219089Spjd{ 1167219089Spjd hdr->b_dva.dva_word[0] = 0; 1168219089Spjd hdr->b_dva.dva_word[1] = 0; 1169219089Spjd hdr->b_birth = 0; 1170219089Spjd} 1171219089Spjd 1172168404Spjdstatic arc_buf_hdr_t * 1173268075Sdelphijbuf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1174168404Spjd{ 1175268075Sdelphij const dva_t *dva = BP_IDENTITY(bp); 1176268075Sdelphij uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1177168404Spjd uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1178168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1179275811Sdelphij arc_buf_hdr_t *hdr; 1180168404Spjd 1181168404Spjd mutex_enter(hash_lock); 1182275811Sdelphij for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1183275811Sdelphij hdr = hdr->b_hash_next) { 1184275811Sdelphij if (BUF_EQUAL(spa, dva, birth, hdr)) { 1185168404Spjd *lockp = hash_lock; 1186275811Sdelphij return (hdr); 1187168404Spjd } 1188168404Spjd } 1189168404Spjd mutex_exit(hash_lock); 1190168404Spjd *lockp = NULL; 1191168404Spjd return (NULL); 1192168404Spjd} 1193168404Spjd 1194168404Spjd/* 1195168404Spjd * Insert an entry into the hash table. If there is already an element 1196168404Spjd * equal to elem in the hash table, then the already existing element 1197168404Spjd * will be returned and the new element will not be inserted. 1198168404Spjd * Otherwise returns NULL. 1199286570Smav * If lockp == NULL, the caller is assumed to already hold the hash lock. 1200168404Spjd */ 1201168404Spjdstatic arc_buf_hdr_t * 1202275811Sdelphijbuf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1203168404Spjd{ 1204275811Sdelphij uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1205168404Spjd kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1206275811Sdelphij arc_buf_hdr_t *fhdr; 1207168404Spjd uint32_t i; 1208168404Spjd 1209275811Sdelphij ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1210275811Sdelphij ASSERT(hdr->b_birth != 0); 1211275811Sdelphij ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1212286570Smav 1213286570Smav if (lockp != NULL) { 1214286570Smav *lockp = hash_lock; 1215286570Smav mutex_enter(hash_lock); 1216286570Smav } else { 1217286570Smav ASSERT(MUTEX_HELD(hash_lock)); 1218286570Smav } 1219286570Smav 1220275811Sdelphij for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1221275811Sdelphij fhdr = fhdr->b_hash_next, i++) { 1222275811Sdelphij if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1223275811Sdelphij return (fhdr); 1224168404Spjd } 1225168404Spjd 1226275811Sdelphij hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1227275811Sdelphij buf_hash_table.ht_table[idx] = hdr; 1228275811Sdelphij hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1229168404Spjd 1230168404Spjd /* collect some hash table performance data */ 1231168404Spjd if (i > 0) { 1232168404Spjd ARCSTAT_BUMP(arcstat_hash_collisions); 1233168404Spjd if (i == 1) 1234168404Spjd ARCSTAT_BUMP(arcstat_hash_chains); 1235168404Spjd 1236168404Spjd ARCSTAT_MAX(arcstat_hash_chain_max, i); 1237168404Spjd } 1238168404Spjd 1239168404Spjd ARCSTAT_BUMP(arcstat_hash_elements); 1240168404Spjd ARCSTAT_MAXSTAT(arcstat_hash_elements); 1241168404Spjd 1242168404Spjd return (NULL); 1243168404Spjd} 1244168404Spjd 1245168404Spjdstatic void 1246275811Sdelphijbuf_hash_remove(arc_buf_hdr_t *hdr) 1247168404Spjd{ 1248275811Sdelphij arc_buf_hdr_t *fhdr, **hdrp; 1249275811Sdelphij uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1250168404Spjd 1251168404Spjd ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1252275811Sdelphij ASSERT(HDR_IN_HASH_TABLE(hdr)); 1253168404Spjd 1254275811Sdelphij hdrp = &buf_hash_table.ht_table[idx]; 1255275811Sdelphij while ((fhdr = *hdrp) != hdr) { 1256275811Sdelphij ASSERT(fhdr != NULL); 1257275811Sdelphij hdrp = &fhdr->b_hash_next; 1258168404Spjd } 1259275811Sdelphij *hdrp = hdr->b_hash_next; 1260275811Sdelphij hdr->b_hash_next = NULL; 1261275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1262168404Spjd 1263168404Spjd /* collect some hash table performance data */ 1264168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1265168404Spjd 1266168404Spjd if (buf_hash_table.ht_table[idx] && 1267168404Spjd buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1268168404Spjd ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1269168404Spjd} 1270168404Spjd 1271168404Spjd/* 1272168404Spjd * Global data structures and functions for the buf kmem cache. 1273168404Spjd */ 1274286570Smavstatic kmem_cache_t *hdr_full_cache; 1275286570Smavstatic kmem_cache_t *hdr_l2only_cache; 1276168404Spjdstatic kmem_cache_t *buf_cache; 1277168404Spjd 1278168404Spjdstatic void 1279168404Spjdbuf_fini(void) 1280168404Spjd{ 1281168404Spjd int i; 1282168404Spjd 1283168404Spjd kmem_free(buf_hash_table.ht_table, 1284168404Spjd (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1285168404Spjd for (i = 0; i < BUF_LOCKS; i++) 1286168404Spjd mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1287286570Smav kmem_cache_destroy(hdr_full_cache); 1288286570Smav kmem_cache_destroy(hdr_l2only_cache); 1289168404Spjd kmem_cache_destroy(buf_cache); 1290168404Spjd} 1291168404Spjd 1292168404Spjd/* 1293168404Spjd * Constructor callback - called when the cache is empty 1294168404Spjd * and a new buf is requested. 1295168404Spjd */ 1296168404Spjd/* ARGSUSED */ 1297168404Spjdstatic int 1298286570Smavhdr_full_cons(void *vbuf, void *unused, int kmflag) 1299168404Spjd{ 1300275811Sdelphij arc_buf_hdr_t *hdr = vbuf; 1301168404Spjd 1302286570Smav bzero(hdr, HDR_FULL_SIZE); 1303286570Smav cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1304286570Smav refcount_create(&hdr->b_l1hdr.b_refcnt); 1305286570Smav mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1306286763Smav multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1307286570Smav arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1308185029Spjd 1309168404Spjd return (0); 1310168404Spjd} 1311168404Spjd 1312185029Spjd/* ARGSUSED */ 1313185029Spjdstatic int 1314286570Smavhdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1315286570Smav{ 1316286570Smav arc_buf_hdr_t *hdr = vbuf; 1317286570Smav 1318286570Smav bzero(hdr, HDR_L2ONLY_SIZE); 1319286570Smav arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1320286570Smav 1321286570Smav return (0); 1322286570Smav} 1323286570Smav 1324286570Smav/* ARGSUSED */ 1325286570Smavstatic int 1326185029Spjdbuf_cons(void *vbuf, void *unused, int kmflag) 1327185029Spjd{ 1328185029Spjd arc_buf_t *buf = vbuf; 1329185029Spjd 1330185029Spjd bzero(buf, sizeof (arc_buf_t)); 1331219089Spjd mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1332208373Smm arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1333208373Smm 1334185029Spjd return (0); 1335185029Spjd} 1336185029Spjd 1337168404Spjd/* 1338168404Spjd * Destructor callback - called when a cached buf is 1339168404Spjd * no longer required. 1340168404Spjd */ 1341168404Spjd/* ARGSUSED */ 1342168404Spjdstatic void 1343286570Smavhdr_full_dest(void *vbuf, void *unused) 1344168404Spjd{ 1345275811Sdelphij arc_buf_hdr_t *hdr = vbuf; 1346168404Spjd 1347275811Sdelphij ASSERT(BUF_EMPTY(hdr)); 1348286570Smav cv_destroy(&hdr->b_l1hdr.b_cv); 1349286570Smav refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1350286570Smav mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1351286763Smav ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1352286570Smav arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1353168404Spjd} 1354168404Spjd 1355185029Spjd/* ARGSUSED */ 1356185029Spjdstatic void 1357286570Smavhdr_l2only_dest(void *vbuf, void *unused) 1358286570Smav{ 1359286570Smav arc_buf_hdr_t *hdr = vbuf; 1360286570Smav 1361286570Smav ASSERT(BUF_EMPTY(hdr)); 1362286570Smav arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1363286570Smav} 1364286570Smav 1365286570Smav/* ARGSUSED */ 1366286570Smavstatic void 1367185029Spjdbuf_dest(void *vbuf, void *unused) 1368185029Spjd{ 1369185029Spjd arc_buf_t *buf = vbuf; 1370185029Spjd 1371219089Spjd mutex_destroy(&buf->b_evict_lock); 1372208373Smm arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1373185029Spjd} 1374185029Spjd 1375168404Spjd/* 1376168404Spjd * Reclaim callback -- invoked when memory is low. 1377168404Spjd */ 1378168404Spjd/* ARGSUSED */ 1379168404Spjdstatic void 1380168404Spjdhdr_recl(void *unused) 1381168404Spjd{ 1382168404Spjd dprintf("hdr_recl called\n"); 1383168404Spjd /* 1384168404Spjd * umem calls the reclaim func when we destroy the buf cache, 1385168404Spjd * which is after we do arc_fini(). 1386168404Spjd */ 1387168404Spjd if (!arc_dead) 1388286763Smav cv_signal(&arc_reclaim_thread_cv); 1389168404Spjd} 1390168404Spjd 1391168404Spjdstatic void 1392168404Spjdbuf_init(void) 1393168404Spjd{ 1394168404Spjd uint64_t *ct; 1395168404Spjd uint64_t hsize = 1ULL << 12; 1396168404Spjd int i, j; 1397168404Spjd 1398168404Spjd /* 1399168404Spjd * The hash table is big enough to fill all of physical memory 1400269230Sdelphij * with an average block size of zfs_arc_average_blocksize (default 8K). 1401269230Sdelphij * By default, the table will take up 1402269230Sdelphij * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1403168404Spjd */ 1404269230Sdelphij while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1405168404Spjd hsize <<= 1; 1406168404Spjdretry: 1407168404Spjd buf_hash_table.ht_mask = hsize - 1; 1408168404Spjd buf_hash_table.ht_table = 1409168404Spjd kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1410168404Spjd if (buf_hash_table.ht_table == NULL) { 1411168404Spjd ASSERT(hsize > (1ULL << 8)); 1412168404Spjd hsize >>= 1; 1413168404Spjd goto retry; 1414168404Spjd } 1415168404Spjd 1416286570Smav hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1417286570Smav 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1418286570Smav hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1419286570Smav HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1420286570Smav NULL, NULL, 0); 1421168404Spjd buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1422185029Spjd 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1423168404Spjd 1424168404Spjd for (i = 0; i < 256; i++) 1425168404Spjd for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1426168404Spjd *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1427168404Spjd 1428168404Spjd for (i = 0; i < BUF_LOCKS; i++) { 1429168404Spjd mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1430168404Spjd NULL, MUTEX_DEFAULT, NULL); 1431168404Spjd } 1432168404Spjd} 1433168404Spjd 1434286570Smav/* 1435286570Smav * Transition between the two allocation states for the arc_buf_hdr struct. 1436286570Smav * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1437286570Smav * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1438286570Smav * version is used when a cache buffer is only in the L2ARC in order to reduce 1439286570Smav * memory usage. 1440286570Smav */ 1441286570Smavstatic arc_buf_hdr_t * 1442286570Smavarc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1443286570Smav{ 1444286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 1445286570Smav 1446286570Smav arc_buf_hdr_t *nhdr; 1447286570Smav l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1448286570Smav 1449286570Smav ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1450286570Smav (old == hdr_l2only_cache && new == hdr_full_cache)); 1451286570Smav 1452286570Smav nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1453286570Smav 1454286570Smav ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1455286570Smav buf_hash_remove(hdr); 1456286570Smav 1457286570Smav bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1458286598Smav 1459286570Smav if (new == hdr_full_cache) { 1460286570Smav nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1461286570Smav /* 1462286570Smav * arc_access and arc_change_state need to be aware that a 1463286570Smav * header has just come out of L2ARC, so we set its state to 1464286570Smav * l2c_only even though it's about to change. 1465286570Smav */ 1466286570Smav nhdr->b_l1hdr.b_state = arc_l2c_only; 1467286763Smav 1468286763Smav /* Verify previous threads set to NULL before freeing */ 1469286763Smav ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1470286570Smav } else { 1471286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 1472286570Smav ASSERT0(hdr->b_l1hdr.b_datacnt); 1473286763Smav 1474286570Smav /* 1475286763Smav * If we've reached here, We must have been called from 1476286763Smav * arc_evict_hdr(), as such we should have already been 1477286763Smav * removed from any ghost list we were previously on 1478286763Smav * (which protects us from racing with arc_evict_state), 1479286763Smav * thus no locking is needed during this check. 1480286570Smav */ 1481286763Smav ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1482286763Smav 1483286763Smav /* 1484286763Smav * A buffer must not be moved into the arc_l2c_only 1485286763Smav * state if it's not finished being written out to the 1486286763Smav * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1487286763Smav * might try to be accessed, even though it was removed. 1488286763Smav */ 1489286763Smav VERIFY(!HDR_L2_WRITING(hdr)); 1490286763Smav VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1491286763Smav 1492286570Smav nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1493286570Smav } 1494286570Smav /* 1495286570Smav * The header has been reallocated so we need to re-insert it into any 1496286570Smav * lists it was on. 1497286570Smav */ 1498286570Smav (void) buf_hash_insert(nhdr, NULL); 1499286570Smav 1500286570Smav ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1501286570Smav 1502286570Smav mutex_enter(&dev->l2ad_mtx); 1503286570Smav 1504286570Smav /* 1505286570Smav * We must place the realloc'ed header back into the list at 1506286570Smav * the same spot. Otherwise, if it's placed earlier in the list, 1507286570Smav * l2arc_write_buffers() could find it during the function's 1508286570Smav * write phase, and try to write it out to the l2arc. 1509286570Smav */ 1510286570Smav list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1511286570Smav list_remove(&dev->l2ad_buflist, hdr); 1512286570Smav 1513286570Smav mutex_exit(&dev->l2ad_mtx); 1514286570Smav 1515286598Smav /* 1516286598Smav * Since we're using the pointer address as the tag when 1517286598Smav * incrementing and decrementing the l2ad_alloc refcount, we 1518286598Smav * must remove the old pointer (that we're about to destroy) and 1519286598Smav * add the new pointer to the refcount. Otherwise we'd remove 1520286598Smav * the wrong pointer address when calling arc_hdr_destroy() later. 1521286598Smav */ 1522286598Smav 1523286598Smav (void) refcount_remove_many(&dev->l2ad_alloc, 1524286598Smav hdr->b_l2hdr.b_asize, hdr); 1525286598Smav 1526286598Smav (void) refcount_add_many(&dev->l2ad_alloc, 1527286598Smav nhdr->b_l2hdr.b_asize, nhdr); 1528286598Smav 1529286570Smav buf_discard_identity(hdr); 1530286570Smav hdr->b_freeze_cksum = NULL; 1531286570Smav kmem_cache_free(old, hdr); 1532286570Smav 1533286570Smav return (nhdr); 1534286570Smav} 1535286570Smav 1536286570Smav 1537168404Spjd#define ARC_MINTIME (hz>>4) /* 62 ms */ 1538168404Spjd 1539168404Spjdstatic void 1540168404Spjdarc_cksum_verify(arc_buf_t *buf) 1541168404Spjd{ 1542168404Spjd zio_cksum_t zc; 1543168404Spjd 1544168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1545168404Spjd return; 1546168404Spjd 1547286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1548286570Smav if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1549286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1550168404Spjd return; 1551168404Spjd } 1552168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1553168404Spjd if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1554168404Spjd panic("buffer modified while frozen!"); 1555286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1556168404Spjd} 1557168404Spjd 1558185029Spjdstatic int 1559185029Spjdarc_cksum_equal(arc_buf_t *buf) 1560185029Spjd{ 1561185029Spjd zio_cksum_t zc; 1562185029Spjd int equal; 1563185029Spjd 1564286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1565185029Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1566185029Spjd equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1567286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1568185029Spjd 1569185029Spjd return (equal); 1570185029Spjd} 1571185029Spjd 1572168404Spjdstatic void 1573185029Spjdarc_cksum_compute(arc_buf_t *buf, boolean_t force) 1574168404Spjd{ 1575185029Spjd if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1576168404Spjd return; 1577168404Spjd 1578286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1579168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1580286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1581168404Spjd return; 1582168404Spjd } 1583168404Spjd buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1584168404Spjd fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1585168404Spjd buf->b_hdr->b_freeze_cksum); 1586286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1587240133Smm#ifdef illumos 1588240133Smm arc_buf_watch(buf); 1589277300Ssmh#endif 1590168404Spjd} 1591168404Spjd 1592240133Smm#ifdef illumos 1593240133Smm#ifndef _KERNEL 1594240133Smmtypedef struct procctl { 1595240133Smm long cmd; 1596240133Smm prwatch_t prwatch; 1597240133Smm} procctl_t; 1598240133Smm#endif 1599240133Smm 1600240133Smm/* ARGSUSED */ 1601240133Smmstatic void 1602240133Smmarc_buf_unwatch(arc_buf_t *buf) 1603240133Smm{ 1604240133Smm#ifndef _KERNEL 1605240133Smm if (arc_watch) { 1606240133Smm int result; 1607240133Smm procctl_t ctl; 1608240133Smm ctl.cmd = PCWATCH; 1609240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1610240133Smm ctl.prwatch.pr_size = 0; 1611240133Smm ctl.prwatch.pr_wflags = 0; 1612240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1613240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1614240133Smm } 1615240133Smm#endif 1616240133Smm} 1617240133Smm 1618240133Smm/* ARGSUSED */ 1619240133Smmstatic void 1620240133Smmarc_buf_watch(arc_buf_t *buf) 1621240133Smm{ 1622240133Smm#ifndef _KERNEL 1623240133Smm if (arc_watch) { 1624240133Smm int result; 1625240133Smm procctl_t ctl; 1626240133Smm ctl.cmd = PCWATCH; 1627240133Smm ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1628240133Smm ctl.prwatch.pr_size = buf->b_hdr->b_size; 1629240133Smm ctl.prwatch.pr_wflags = WA_WRITE; 1630240133Smm result = write(arc_procfd, &ctl, sizeof (ctl)); 1631240133Smm ASSERT3U(result, ==, sizeof (ctl)); 1632240133Smm } 1633240133Smm#endif 1634240133Smm} 1635240133Smm#endif /* illumos */ 1636240133Smm 1637286570Smavstatic arc_buf_contents_t 1638286570Smavarc_buf_type(arc_buf_hdr_t *hdr) 1639286570Smav{ 1640286570Smav if (HDR_ISTYPE_METADATA(hdr)) { 1641286570Smav return (ARC_BUFC_METADATA); 1642286570Smav } else { 1643286570Smav return (ARC_BUFC_DATA); 1644286570Smav } 1645286570Smav} 1646286570Smav 1647286570Smavstatic uint32_t 1648286570Smavarc_bufc_to_flags(arc_buf_contents_t type) 1649286570Smav{ 1650286570Smav switch (type) { 1651286570Smav case ARC_BUFC_DATA: 1652286570Smav /* metadata field is 0 if buffer contains normal data */ 1653286570Smav return (0); 1654286570Smav case ARC_BUFC_METADATA: 1655286570Smav return (ARC_FLAG_BUFC_METADATA); 1656286570Smav default: 1657286570Smav break; 1658286570Smav } 1659286570Smav panic("undefined ARC buffer type!"); 1660286570Smav return ((uint32_t)-1); 1661286570Smav} 1662286570Smav 1663168404Spjdvoid 1664168404Spjdarc_buf_thaw(arc_buf_t *buf) 1665168404Spjd{ 1666185029Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1667286570Smav if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1668185029Spjd panic("modifying non-anon buffer!"); 1669286570Smav if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1670185029Spjd panic("modifying buffer while i/o in progress!"); 1671185029Spjd arc_cksum_verify(buf); 1672185029Spjd } 1673168404Spjd 1674286570Smav mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1675168404Spjd if (buf->b_hdr->b_freeze_cksum != NULL) { 1676168404Spjd kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1677168404Spjd buf->b_hdr->b_freeze_cksum = NULL; 1678168404Spjd } 1679219089Spjd 1680286570Smav#ifdef ZFS_DEBUG 1681219089Spjd if (zfs_flags & ZFS_DEBUG_MODIFY) { 1682286570Smav if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1683286570Smav kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1684286570Smav buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1685219089Spjd } 1686286570Smav#endif 1687219089Spjd 1688286570Smav mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1689240133Smm 1690240133Smm#ifdef illumos 1691240133Smm arc_buf_unwatch(buf); 1692277300Ssmh#endif 1693168404Spjd} 1694168404Spjd 1695168404Spjdvoid 1696168404Spjdarc_buf_freeze(arc_buf_t *buf) 1697168404Spjd{ 1698219089Spjd kmutex_t *hash_lock; 1699219089Spjd 1700168404Spjd if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1701168404Spjd return; 1702168404Spjd 1703219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 1704219089Spjd mutex_enter(hash_lock); 1705219089Spjd 1706168404Spjd ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1707286570Smav buf->b_hdr->b_l1hdr.b_state == arc_anon); 1708185029Spjd arc_cksum_compute(buf, B_FALSE); 1709219089Spjd mutex_exit(hash_lock); 1710240133Smm 1711168404Spjd} 1712168404Spjd 1713168404Spjdstatic void 1714275811Sdelphijadd_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1715168404Spjd{ 1716286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1717168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1718286570Smav arc_state_t *state = hdr->b_l1hdr.b_state; 1719168404Spjd 1720286570Smav if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1721286570Smav (state != arc_anon)) { 1722286570Smav /* We don't use the L2-only state list. */ 1723286570Smav if (state != arc_l2c_only) { 1724286763Smav arc_buf_contents_t type = arc_buf_type(hdr); 1725286570Smav uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1726286763Smav multilist_t *list = &state->arcs_list[type]; 1727286763Smav uint64_t *size = &state->arcs_lsize[type]; 1728168404Spjd 1729286763Smav multilist_remove(list, hdr); 1730286763Smav 1731286570Smav if (GHOST_STATE(state)) { 1732286570Smav ASSERT0(hdr->b_l1hdr.b_datacnt); 1733286570Smav ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1734286570Smav delta = hdr->b_size; 1735286570Smav } 1736286570Smav ASSERT(delta > 0); 1737286570Smav ASSERT3U(*size, >=, delta); 1738286570Smav atomic_add_64(size, -delta); 1739168404Spjd } 1740185029Spjd /* remove the prefetch flag if we get a reference */ 1741286570Smav hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1742168404Spjd } 1743168404Spjd} 1744168404Spjd 1745168404Spjdstatic int 1746275811Sdelphijremove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1747168404Spjd{ 1748168404Spjd int cnt; 1749286570Smav arc_state_t *state = hdr->b_l1hdr.b_state; 1750168404Spjd 1751286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1752168404Spjd ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1753168404Spjd ASSERT(!GHOST_STATE(state)); 1754168404Spjd 1755286570Smav /* 1756286570Smav * arc_l2c_only counts as a ghost state so we don't need to explicitly 1757286570Smav * check to prevent usage of the arc_l2c_only list. 1758286570Smav */ 1759286570Smav if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1760168404Spjd (state != arc_anon)) { 1761286763Smav arc_buf_contents_t type = arc_buf_type(hdr); 1762286763Smav multilist_t *list = &state->arcs_list[type]; 1763286763Smav uint64_t *size = &state->arcs_lsize[type]; 1764185029Spjd 1765286763Smav multilist_insert(list, hdr); 1766286763Smav 1767286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1768286570Smav atomic_add_64(size, hdr->b_size * 1769286570Smav hdr->b_l1hdr.b_datacnt); 1770168404Spjd } 1771168404Spjd return (cnt); 1772168404Spjd} 1773168404Spjd 1774168404Spjd/* 1775286763Smav * Move the supplied buffer to the indicated state. The hash lock 1776168404Spjd * for the buffer must be held by the caller. 1777168404Spjd */ 1778168404Spjdstatic void 1779275811Sdelphijarc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1780275811Sdelphij kmutex_t *hash_lock) 1781168404Spjd{ 1782286570Smav arc_state_t *old_state; 1783286570Smav int64_t refcnt; 1784286570Smav uint32_t datacnt; 1785168404Spjd uint64_t from_delta, to_delta; 1786286570Smav arc_buf_contents_t buftype = arc_buf_type(hdr); 1787168404Spjd 1788286570Smav /* 1789286570Smav * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1790286570Smav * in arc_read() when bringing a buffer out of the L2ARC. However, the 1791286570Smav * L1 hdr doesn't always exist when we change state to arc_anon before 1792286570Smav * destroying a header, in which case reallocating to add the L1 hdr is 1793286570Smav * pointless. 1794286570Smav */ 1795286570Smav if (HDR_HAS_L1HDR(hdr)) { 1796286570Smav old_state = hdr->b_l1hdr.b_state; 1797286570Smav refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1798286570Smav datacnt = hdr->b_l1hdr.b_datacnt; 1799286570Smav } else { 1800286570Smav old_state = arc_l2c_only; 1801286570Smav refcnt = 0; 1802286570Smav datacnt = 0; 1803286570Smav } 1804286570Smav 1805168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 1806258632Savg ASSERT3P(new_state, !=, old_state); 1807286570Smav ASSERT(refcnt == 0 || datacnt > 0); 1808286570Smav ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1809286570Smav ASSERT(old_state != arc_anon || datacnt <= 1); 1810168404Spjd 1811286570Smav from_delta = to_delta = datacnt * hdr->b_size; 1812168404Spjd 1813168404Spjd /* 1814168404Spjd * If this buffer is evictable, transfer it from the 1815168404Spjd * old state list to the new state list. 1816168404Spjd */ 1817168404Spjd if (refcnt == 0) { 1818286570Smav if (old_state != arc_anon && old_state != arc_l2c_only) { 1819286570Smav uint64_t *size = &old_state->arcs_lsize[buftype]; 1820168404Spjd 1821286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1822286763Smav multilist_remove(&old_state->arcs_list[buftype], hdr); 1823168404Spjd 1824168404Spjd /* 1825168404Spjd * If prefetching out of the ghost cache, 1826219089Spjd * we will have a non-zero datacnt. 1827168404Spjd */ 1828286570Smav if (GHOST_STATE(old_state) && datacnt == 0) { 1829168404Spjd /* ghost elements have a ghost size */ 1830286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 1831275811Sdelphij from_delta = hdr->b_size; 1832168404Spjd } 1833185029Spjd ASSERT3U(*size, >=, from_delta); 1834185029Spjd atomic_add_64(size, -from_delta); 1835168404Spjd } 1836286570Smav if (new_state != arc_anon && new_state != arc_l2c_only) { 1837286570Smav uint64_t *size = &new_state->arcs_lsize[buftype]; 1838168404Spjd 1839286570Smav /* 1840286570Smav * An L1 header always exists here, since if we're 1841286570Smav * moving to some L1-cached state (i.e. not l2c_only or 1842286570Smav * anonymous), we realloc the header to add an L1hdr 1843286570Smav * beforehand. 1844286570Smav */ 1845286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 1846286763Smav multilist_insert(&new_state->arcs_list[buftype], hdr); 1847168404Spjd 1848168404Spjd /* ghost elements have a ghost size */ 1849168404Spjd if (GHOST_STATE(new_state)) { 1850286762Smav ASSERT0(datacnt); 1851286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 1852275811Sdelphij to_delta = hdr->b_size; 1853168404Spjd } 1854185029Spjd atomic_add_64(size, to_delta); 1855168404Spjd } 1856168404Spjd } 1857168404Spjd 1858275811Sdelphij ASSERT(!BUF_EMPTY(hdr)); 1859275811Sdelphij if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1860275811Sdelphij buf_hash_remove(hdr); 1861168404Spjd 1862286570Smav /* adjust state sizes (ignore arc_l2c_only) */ 1863286570Smav if (to_delta && new_state != arc_l2c_only) 1864168404Spjd atomic_add_64(&new_state->arcs_size, to_delta); 1865286570Smav if (from_delta && old_state != arc_l2c_only) { 1866168404Spjd ASSERT3U(old_state->arcs_size, >=, from_delta); 1867168404Spjd atomic_add_64(&old_state->arcs_size, -from_delta); 1868168404Spjd } 1869286570Smav if (HDR_HAS_L1HDR(hdr)) 1870286570Smav hdr->b_l1hdr.b_state = new_state; 1871185029Spjd 1872286570Smav /* 1873286570Smav * L2 headers should never be on the L2 state list since they don't 1874286570Smav * have L1 headers allocated. 1875286570Smav */ 1876286763Smav ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1877286763Smav multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1878168404Spjd} 1879168404Spjd 1880185029Spjdvoid 1881208373Smmarc_space_consume(uint64_t space, arc_space_type_t type) 1882185029Spjd{ 1883208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1884208373Smm 1885208373Smm switch (type) { 1886208373Smm case ARC_SPACE_DATA: 1887208373Smm ARCSTAT_INCR(arcstat_data_size, space); 1888208373Smm break; 1889286574Smav case ARC_SPACE_META: 1890286574Smav ARCSTAT_INCR(arcstat_metadata_size, space); 1891286574Smav break; 1892208373Smm case ARC_SPACE_OTHER: 1893208373Smm ARCSTAT_INCR(arcstat_other_size, space); 1894208373Smm break; 1895208373Smm case ARC_SPACE_HDRS: 1896208373Smm ARCSTAT_INCR(arcstat_hdr_size, space); 1897208373Smm break; 1898208373Smm case ARC_SPACE_L2HDRS: 1899208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1900208373Smm break; 1901208373Smm } 1902208373Smm 1903286574Smav if (type != ARC_SPACE_DATA) 1904286574Smav ARCSTAT_INCR(arcstat_meta_used, space); 1905286574Smav 1906185029Spjd atomic_add_64(&arc_size, space); 1907185029Spjd} 1908185029Spjd 1909185029Spjdvoid 1910208373Smmarc_space_return(uint64_t space, arc_space_type_t type) 1911185029Spjd{ 1912208373Smm ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1913208373Smm 1914208373Smm switch (type) { 1915208373Smm case ARC_SPACE_DATA: 1916208373Smm ARCSTAT_INCR(arcstat_data_size, -space); 1917208373Smm break; 1918286574Smav case ARC_SPACE_META: 1919286574Smav ARCSTAT_INCR(arcstat_metadata_size, -space); 1920286574Smav break; 1921208373Smm case ARC_SPACE_OTHER: 1922208373Smm ARCSTAT_INCR(arcstat_other_size, -space); 1923208373Smm break; 1924208373Smm case ARC_SPACE_HDRS: 1925208373Smm ARCSTAT_INCR(arcstat_hdr_size, -space); 1926208373Smm break; 1927208373Smm case ARC_SPACE_L2HDRS: 1928208373Smm ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1929208373Smm break; 1930208373Smm } 1931208373Smm 1932286574Smav if (type != ARC_SPACE_DATA) { 1933286574Smav ASSERT(arc_meta_used >= space); 1934286574Smav if (arc_meta_max < arc_meta_used) 1935286574Smav arc_meta_max = arc_meta_used; 1936286574Smav ARCSTAT_INCR(arcstat_meta_used, -space); 1937286574Smav } 1938286574Smav 1939185029Spjd ASSERT(arc_size >= space); 1940185029Spjd atomic_add_64(&arc_size, -space); 1941185029Spjd} 1942185029Spjd 1943168404Spjdarc_buf_t * 1944286570Smavarc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 1945168404Spjd{ 1946168404Spjd arc_buf_hdr_t *hdr; 1947168404Spjd arc_buf_t *buf; 1948168404Spjd 1949168404Spjd ASSERT3U(size, >, 0); 1950286570Smav hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 1951168404Spjd ASSERT(BUF_EMPTY(hdr)); 1952286570Smav ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 1953168404Spjd hdr->b_size = size; 1954228103Smm hdr->b_spa = spa_load_guid(spa); 1955286570Smav 1956185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1957168404Spjd buf->b_hdr = hdr; 1958168404Spjd buf->b_data = NULL; 1959168404Spjd buf->b_efunc = NULL; 1960168404Spjd buf->b_private = NULL; 1961168404Spjd buf->b_next = NULL; 1962286570Smav 1963286570Smav hdr->b_flags = arc_bufc_to_flags(type); 1964286570Smav hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1965286570Smav 1966286570Smav hdr->b_l1hdr.b_buf = buf; 1967286570Smav hdr->b_l1hdr.b_state = arc_anon; 1968286570Smav hdr->b_l1hdr.b_arc_access = 0; 1969286570Smav hdr->b_l1hdr.b_datacnt = 1; 1970286763Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 1971286570Smav 1972168404Spjd arc_get_data_buf(buf); 1973286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 1974286570Smav (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1975168404Spjd 1976168404Spjd return (buf); 1977168404Spjd} 1978168404Spjd 1979209962Smmstatic char *arc_onloan_tag = "onloan"; 1980209962Smm 1981209962Smm/* 1982209962Smm * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1983209962Smm * flight data by arc_tempreserve_space() until they are "returned". Loaned 1984209962Smm * buffers must be returned to the arc before they can be used by the DMU or 1985209962Smm * freed. 1986209962Smm */ 1987209962Smmarc_buf_t * 1988209962Smmarc_loan_buf(spa_t *spa, int size) 1989209962Smm{ 1990209962Smm arc_buf_t *buf; 1991209962Smm 1992209962Smm buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1993209962Smm 1994209962Smm atomic_add_64(&arc_loaned_bytes, size); 1995209962Smm return (buf); 1996209962Smm} 1997209962Smm 1998209962Smm/* 1999209962Smm * Return a loaned arc buffer to the arc. 2000209962Smm */ 2001209962Smmvoid 2002209962Smmarc_return_buf(arc_buf_t *buf, void *tag) 2003209962Smm{ 2004209962Smm arc_buf_hdr_t *hdr = buf->b_hdr; 2005209962Smm 2006209962Smm ASSERT(buf->b_data != NULL); 2007286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 2008286570Smav (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2009286570Smav (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2010209962Smm 2011209962Smm atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2012209962Smm} 2013209962Smm 2014219089Spjd/* Detach an arc_buf from a dbuf (tag) */ 2015219089Spjdvoid 2016219089Spjdarc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2017219089Spjd{ 2018286570Smav arc_buf_hdr_t *hdr = buf->b_hdr; 2019219089Spjd 2020219089Spjd ASSERT(buf->b_data != NULL); 2021286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 2022286570Smav (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2023286570Smav (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2024219089Spjd buf->b_efunc = NULL; 2025219089Spjd buf->b_private = NULL; 2026219089Spjd 2027219089Spjd atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2028219089Spjd} 2029219089Spjd 2030168404Spjdstatic arc_buf_t * 2031168404Spjdarc_buf_clone(arc_buf_t *from) 2032168404Spjd{ 2033168404Spjd arc_buf_t *buf; 2034168404Spjd arc_buf_hdr_t *hdr = from->b_hdr; 2035168404Spjd uint64_t size = hdr->b_size; 2036168404Spjd 2037286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 2038286570Smav ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2039219089Spjd 2040185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2041168404Spjd buf->b_hdr = hdr; 2042168404Spjd buf->b_data = NULL; 2043168404Spjd buf->b_efunc = NULL; 2044168404Spjd buf->b_private = NULL; 2045286570Smav buf->b_next = hdr->b_l1hdr.b_buf; 2046286570Smav hdr->b_l1hdr.b_buf = buf; 2047168404Spjd arc_get_data_buf(buf); 2048168404Spjd bcopy(from->b_data, buf->b_data, size); 2049242845Sdelphij 2050242845Sdelphij /* 2051242845Sdelphij * This buffer already exists in the arc so create a duplicate 2052242845Sdelphij * copy for the caller. If the buffer is associated with user data 2053242845Sdelphij * then track the size and number of duplicates. These stats will be 2054242845Sdelphij * updated as duplicate buffers are created and destroyed. 2055242845Sdelphij */ 2056286570Smav if (HDR_ISTYPE_DATA(hdr)) { 2057242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_buffers); 2058242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2059242845Sdelphij } 2060286570Smav hdr->b_l1hdr.b_datacnt += 1; 2061168404Spjd return (buf); 2062168404Spjd} 2063168404Spjd 2064168404Spjdvoid 2065168404Spjdarc_buf_add_ref(arc_buf_t *buf, void* tag) 2066168404Spjd{ 2067168404Spjd arc_buf_hdr_t *hdr; 2068168404Spjd kmutex_t *hash_lock; 2069168404Spjd 2070168404Spjd /* 2071185029Spjd * Check to see if this buffer is evicted. Callers 2072185029Spjd * must verify b_data != NULL to know if the add_ref 2073185029Spjd * was successful. 2074168404Spjd */ 2075219089Spjd mutex_enter(&buf->b_evict_lock); 2076185029Spjd if (buf->b_data == NULL) { 2077219089Spjd mutex_exit(&buf->b_evict_lock); 2078168404Spjd return; 2079168404Spjd } 2080219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 2081219089Spjd mutex_enter(hash_lock); 2082185029Spjd hdr = buf->b_hdr; 2083286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 2084219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2085219089Spjd mutex_exit(&buf->b_evict_lock); 2086168404Spjd 2087286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2088286570Smav hdr->b_l1hdr.b_state == arc_mfu); 2089286570Smav 2090168404Spjd add_reference(hdr, hash_lock, tag); 2091208373Smm DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2092168404Spjd arc_access(hdr, hash_lock); 2093168404Spjd mutex_exit(hash_lock); 2094168404Spjd ARCSTAT_BUMP(arcstat_hits); 2095286570Smav ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2096286570Smav demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2097168404Spjd data, metadata, hits); 2098168404Spjd} 2099168404Spjd 2100274172Savgstatic void 2101274172Savgarc_buf_free_on_write(void *data, size_t size, 2102274172Savg void (*free_func)(void *, size_t)) 2103274172Savg{ 2104274172Savg l2arc_data_free_t *df; 2105274172Savg 2106286763Smav df = kmem_alloc(sizeof (*df), KM_SLEEP); 2107274172Savg df->l2df_data = data; 2108274172Savg df->l2df_size = size; 2109274172Savg df->l2df_func = free_func; 2110274172Savg mutex_enter(&l2arc_free_on_write_mtx); 2111274172Savg list_insert_head(l2arc_free_on_write, df); 2112274172Savg mutex_exit(&l2arc_free_on_write_mtx); 2113274172Savg} 2114274172Savg 2115185029Spjd/* 2116185029Spjd * Free the arc data buffer. If it is an l2arc write in progress, 2117185029Spjd * the buffer is placed on l2arc_free_on_write to be freed later. 2118185029Spjd */ 2119168404Spjdstatic void 2120240133Smmarc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2121185029Spjd{ 2122240133Smm arc_buf_hdr_t *hdr = buf->b_hdr; 2123240133Smm 2124185029Spjd if (HDR_L2_WRITING(hdr)) { 2125274172Savg arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2126185029Spjd ARCSTAT_BUMP(arcstat_l2_free_on_write); 2127185029Spjd } else { 2128240133Smm free_func(buf->b_data, hdr->b_size); 2129185029Spjd } 2130185029Spjd} 2131185029Spjd 2132268858Sdelphij/* 2133268858Sdelphij * Free up buf->b_data and if 'remove' is set, then pull the 2134268858Sdelphij * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2135268858Sdelphij */ 2136185029Spjdstatic void 2137274172Savgarc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2138274172Savg{ 2139286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 2140286570Smav ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2141274172Savg 2142286570Smav /* 2143286570Smav * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2144286570Smav * that doesn't exist, the header is in the arc_l2c_only state, 2145286570Smav * and there isn't anything to free (it's already been freed). 2146286570Smav */ 2147286570Smav if (!HDR_HAS_L1HDR(hdr)) 2148286570Smav return; 2149274172Savg 2150286763Smav /* 2151286763Smav * The header isn't being written to the l2arc device, thus it 2152286763Smav * shouldn't have a b_tmp_cdata to free. 2153286763Smav */ 2154286763Smav if (!HDR_L2_WRITING(hdr)) { 2155286763Smav ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2156274172Savg return; 2157286763Smav } 2158274172Savg 2159286763Smav /* 2160286763Smav * The header does not have compression enabled. This can be due 2161286763Smav * to the buffer not being compressible, or because we're 2162286763Smav * freeing the buffer before the second phase of 2163286763Smav * l2arc_write_buffer() has started (which does the compression 2164286763Smav * step). In either case, b_tmp_cdata does not point to a 2165286763Smav * separately compressed buffer, so there's nothing to free (it 2166286763Smav * points to the same buffer as the arc_buf_t's b_data field). 2167286763Smav */ 2168286763Smav if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) { 2169286763Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 2170286763Smav return; 2171286763Smav } 2172286570Smav 2173286763Smav /* 2174286763Smav * There's nothing to free since the buffer was all zero's and 2175286763Smav * compressed to a zero length buffer. 2176286763Smav */ 2177286763Smav if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) { 2178286763Smav ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2179286763Smav return; 2180286763Smav } 2181286763Smav 2182286763Smav ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))); 2183286763Smav 2184286763Smav arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2185286763Smav hdr->b_size, zio_data_buf_free); 2186286763Smav 2187274172Savg ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2188286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 2189274172Savg} 2190274172Savg 2191274172Savgstatic void 2192286763Smavarc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2193168404Spjd{ 2194168404Spjd arc_buf_t **bufp; 2195168404Spjd 2196168404Spjd /* free up data associated with the buf */ 2197286570Smav if (buf->b_data != NULL) { 2198286570Smav arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2199168404Spjd uint64_t size = buf->b_hdr->b_size; 2200286570Smav arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2201168404Spjd 2202168404Spjd arc_cksum_verify(buf); 2203240133Smm#ifdef illumos 2204240133Smm arc_buf_unwatch(buf); 2205277300Ssmh#endif 2206219089Spjd 2207286763Smav if (type == ARC_BUFC_METADATA) { 2208286763Smav arc_buf_data_free(buf, zio_buf_free); 2209286763Smav arc_space_return(size, ARC_SPACE_META); 2210286763Smav } else { 2211286763Smav ASSERT(type == ARC_BUFC_DATA); 2212286763Smav arc_buf_data_free(buf, zio_data_buf_free); 2213286763Smav arc_space_return(size, ARC_SPACE_DATA); 2214168404Spjd } 2215286763Smav 2216286763Smav /* protected by hash lock, if in the hash table */ 2217286763Smav if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2218185029Spjd uint64_t *cnt = &state->arcs_lsize[type]; 2219185029Spjd 2220286570Smav ASSERT(refcount_is_zero( 2221286570Smav &buf->b_hdr->b_l1hdr.b_refcnt)); 2222286570Smav ASSERT(state != arc_anon && state != arc_l2c_only); 2223185029Spjd 2224185029Spjd ASSERT3U(*cnt, >=, size); 2225185029Spjd atomic_add_64(cnt, -size); 2226168404Spjd } 2227168404Spjd ASSERT3U(state->arcs_size, >=, size); 2228168404Spjd atomic_add_64(&state->arcs_size, -size); 2229168404Spjd buf->b_data = NULL; 2230242845Sdelphij 2231242845Sdelphij /* 2232242845Sdelphij * If we're destroying a duplicate buffer make sure 2233242845Sdelphij * that the appropriate statistics are updated. 2234242845Sdelphij */ 2235286570Smav if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2236286570Smav HDR_ISTYPE_DATA(buf->b_hdr)) { 2237242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2238242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2239242845Sdelphij } 2240286570Smav ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2241286570Smav buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2242168404Spjd } 2243168404Spjd 2244168404Spjd /* only remove the buf if requested */ 2245268858Sdelphij if (!remove) 2246168404Spjd return; 2247168404Spjd 2248168404Spjd /* remove the buf from the hdr list */ 2249286570Smav for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2250286570Smav bufp = &(*bufp)->b_next) 2251168404Spjd continue; 2252168404Spjd *bufp = buf->b_next; 2253219089Spjd buf->b_next = NULL; 2254168404Spjd 2255168404Spjd ASSERT(buf->b_efunc == NULL); 2256168404Spjd 2257168404Spjd /* clean up the buf */ 2258168404Spjd buf->b_hdr = NULL; 2259168404Spjd kmem_cache_free(buf_cache, buf); 2260168404Spjd} 2261168404Spjd 2262168404Spjdstatic void 2263286598Smavarc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2264286598Smav{ 2265286598Smav l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2266286598Smav l2arc_dev_t *dev = l2hdr->b_dev; 2267286598Smav 2268286598Smav ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2269286598Smav ASSERT(HDR_HAS_L2HDR(hdr)); 2270286598Smav 2271286598Smav list_remove(&dev->l2ad_buflist, hdr); 2272286598Smav 2273286598Smav /* 2274286598Smav * We don't want to leak the b_tmp_cdata buffer that was 2275286598Smav * allocated in l2arc_write_buffers() 2276286598Smav */ 2277286598Smav arc_buf_l2_cdata_free(hdr); 2278286598Smav 2279286598Smav /* 2280286598Smav * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2281286598Smav * this header is being processed by l2arc_write_buffers() (i.e. 2282286598Smav * it's in the first stage of l2arc_write_buffers()). 2283286598Smav * Re-affirming that truth here, just to serve as a reminder. If 2284286598Smav * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2285286598Smav * may not have its HDR_L2_WRITING flag set. (the write may have 2286286598Smav * completed, in which case HDR_L2_WRITING will be false and the 2287286598Smav * b_daddr field will point to the address of the buffer on disk). 2288286598Smav */ 2289286598Smav IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2290286598Smav 2291286598Smav /* 2292286598Smav * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2293286598Smav * l2arc_write_buffers(). Since we've just removed this header 2294286598Smav * from the l2arc buffer list, this header will never reach the 2295286598Smav * second stage of l2arc_write_buffers(), which increments the 2296286598Smav * accounting stats for this header. Thus, we must be careful 2297286598Smav * not to decrement them for this header either. 2298286598Smav */ 2299286598Smav if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2300286598Smav ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2301286598Smav ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2302286598Smav 2303286598Smav vdev_space_update(dev->l2ad_vdev, 2304286598Smav -l2hdr->b_asize, 0, 0); 2305286598Smav 2306286598Smav (void) refcount_remove_many(&dev->l2ad_alloc, 2307286598Smav l2hdr->b_asize, hdr); 2308286598Smav } 2309286598Smav 2310286598Smav hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2311286598Smav} 2312286598Smav 2313286598Smavstatic void 2314168404Spjdarc_hdr_destroy(arc_buf_hdr_t *hdr) 2315168404Spjd{ 2316286570Smav if (HDR_HAS_L1HDR(hdr)) { 2317286570Smav ASSERT(hdr->b_l1hdr.b_buf == NULL || 2318286570Smav hdr->b_l1hdr.b_datacnt > 0); 2319286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2320286570Smav ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2321286570Smav } 2322168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2323286570Smav ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2324168404Spjd 2325286570Smav if (HDR_HAS_L2HDR(hdr)) { 2326286598Smav l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2327286598Smav boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2328286570Smav 2329286598Smav if (!buflist_held) 2330286598Smav mutex_enter(&dev->l2ad_mtx); 2331219089Spjd 2332286570Smav /* 2333286598Smav * Even though we checked this conditional above, we 2334286598Smav * need to check this again now that we have the 2335286598Smav * l2ad_mtx. This is because we could be racing with 2336286598Smav * another thread calling l2arc_evict() which might have 2337286598Smav * destroyed this header's L2 portion as we were waiting 2338286598Smav * to acquire the l2ad_mtx. If that happens, we don't 2339286598Smav * want to re-destroy the header's L2 portion. 2340286570Smav */ 2341286598Smav if (HDR_HAS_L2HDR(hdr)) { 2342286647Smav if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2343286647Smav trim_map_free(dev->l2ad_vdev, 2344286647Smav hdr->b_l2hdr.b_daddr, 2345286647Smav hdr->b_l2hdr.b_asize, 0); 2346286598Smav arc_hdr_l2hdr_destroy(hdr); 2347286598Smav } 2348286570Smav 2349219089Spjd if (!buflist_held) 2350286598Smav mutex_exit(&dev->l2ad_mtx); 2351185029Spjd } 2352185029Spjd 2353286570Smav if (!BUF_EMPTY(hdr)) 2354219089Spjd buf_discard_identity(hdr); 2355168404Spjd if (hdr->b_freeze_cksum != NULL) { 2356168404Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2357168404Spjd hdr->b_freeze_cksum = NULL; 2358168404Spjd } 2359286570Smav 2360286570Smav if (HDR_HAS_L1HDR(hdr)) { 2361286570Smav while (hdr->b_l1hdr.b_buf) { 2362286570Smav arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2363286570Smav 2364286570Smav if (buf->b_efunc != NULL) { 2365286763Smav mutex_enter(&arc_user_evicts_lock); 2366286570Smav mutex_enter(&buf->b_evict_lock); 2367286570Smav ASSERT(buf->b_hdr != NULL); 2368286763Smav arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2369286570Smav hdr->b_l1hdr.b_buf = buf->b_next; 2370286570Smav buf->b_hdr = &arc_eviction_hdr; 2371286570Smav buf->b_next = arc_eviction_list; 2372286570Smav arc_eviction_list = buf; 2373286570Smav mutex_exit(&buf->b_evict_lock); 2374286763Smav cv_signal(&arc_user_evicts_cv); 2375286763Smav mutex_exit(&arc_user_evicts_lock); 2376286570Smav } else { 2377286763Smav arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2378286570Smav } 2379286570Smav } 2380286570Smav#ifdef ZFS_DEBUG 2381286570Smav if (hdr->b_l1hdr.b_thawed != NULL) { 2382286570Smav kmem_free(hdr->b_l1hdr.b_thawed, 1); 2383286570Smav hdr->b_l1hdr.b_thawed = NULL; 2384286570Smav } 2385286570Smav#endif 2386219089Spjd } 2387168404Spjd 2388168404Spjd ASSERT3P(hdr->b_hash_next, ==, NULL); 2389286570Smav if (HDR_HAS_L1HDR(hdr)) { 2390286763Smav ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2391286570Smav ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2392286570Smav kmem_cache_free(hdr_full_cache, hdr); 2393286570Smav } else { 2394286570Smav kmem_cache_free(hdr_l2only_cache, hdr); 2395286570Smav } 2396168404Spjd} 2397168404Spjd 2398168404Spjdvoid 2399168404Spjdarc_buf_free(arc_buf_t *buf, void *tag) 2400168404Spjd{ 2401168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2402286570Smav int hashed = hdr->b_l1hdr.b_state != arc_anon; 2403168404Spjd 2404168404Spjd ASSERT(buf->b_efunc == NULL); 2405168404Spjd ASSERT(buf->b_data != NULL); 2406168404Spjd 2407168404Spjd if (hashed) { 2408168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 2409168404Spjd 2410168404Spjd mutex_enter(hash_lock); 2411219089Spjd hdr = buf->b_hdr; 2412219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2413219089Spjd 2414168404Spjd (void) remove_reference(hdr, hash_lock, tag); 2415286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 2416286763Smav arc_buf_destroy(buf, TRUE); 2417219089Spjd } else { 2418286570Smav ASSERT(buf == hdr->b_l1hdr.b_buf); 2419219089Spjd ASSERT(buf->b_efunc == NULL); 2420275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2421219089Spjd } 2422168404Spjd mutex_exit(hash_lock); 2423168404Spjd } else if (HDR_IO_IN_PROGRESS(hdr)) { 2424168404Spjd int destroy_hdr; 2425168404Spjd /* 2426168404Spjd * We are in the middle of an async write. Don't destroy 2427168404Spjd * this buffer unless the write completes before we finish 2428168404Spjd * decrementing the reference count. 2429168404Spjd */ 2430286763Smav mutex_enter(&arc_user_evicts_lock); 2431168404Spjd (void) remove_reference(hdr, NULL, tag); 2432286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2433168404Spjd destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2434286763Smav mutex_exit(&arc_user_evicts_lock); 2435168404Spjd if (destroy_hdr) 2436168404Spjd arc_hdr_destroy(hdr); 2437168404Spjd } else { 2438219089Spjd if (remove_reference(hdr, NULL, tag) > 0) 2439286763Smav arc_buf_destroy(buf, TRUE); 2440219089Spjd else 2441168404Spjd arc_hdr_destroy(hdr); 2442168404Spjd } 2443168404Spjd} 2444168404Spjd 2445248571Smmboolean_t 2446168404Spjdarc_buf_remove_ref(arc_buf_t *buf, void* tag) 2447168404Spjd{ 2448168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 2449168404Spjd kmutex_t *hash_lock = HDR_LOCK(hdr); 2450248571Smm boolean_t no_callback = (buf->b_efunc == NULL); 2451168404Spjd 2452286570Smav if (hdr->b_l1hdr.b_state == arc_anon) { 2453286570Smav ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2454168404Spjd arc_buf_free(buf, tag); 2455168404Spjd return (no_callback); 2456168404Spjd } 2457168404Spjd 2458168404Spjd mutex_enter(hash_lock); 2459219089Spjd hdr = buf->b_hdr; 2460286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2461219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2462286570Smav ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2463168404Spjd ASSERT(buf->b_data != NULL); 2464168404Spjd 2465168404Spjd (void) remove_reference(hdr, hash_lock, tag); 2466286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 2467168404Spjd if (no_callback) 2468286763Smav arc_buf_destroy(buf, TRUE); 2469168404Spjd } else if (no_callback) { 2470286570Smav ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2471219089Spjd ASSERT(buf->b_efunc == NULL); 2472275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2473168404Spjd } 2474286570Smav ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2475286570Smav refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2476168404Spjd mutex_exit(hash_lock); 2477168404Spjd return (no_callback); 2478168404Spjd} 2479168404Spjd 2480286570Smavint32_t 2481168404Spjdarc_buf_size(arc_buf_t *buf) 2482168404Spjd{ 2483168404Spjd return (buf->b_hdr->b_size); 2484168404Spjd} 2485168404Spjd 2486168404Spjd/* 2487242845Sdelphij * Called from the DMU to determine if the current buffer should be 2488242845Sdelphij * evicted. In order to ensure proper locking, the eviction must be initiated 2489242845Sdelphij * from the DMU. Return true if the buffer is associated with user data and 2490242845Sdelphij * duplicate buffers still exist. 2491242845Sdelphij */ 2492242845Sdelphijboolean_t 2493242845Sdelphijarc_buf_eviction_needed(arc_buf_t *buf) 2494242845Sdelphij{ 2495242845Sdelphij arc_buf_hdr_t *hdr; 2496242845Sdelphij boolean_t evict_needed = B_FALSE; 2497242845Sdelphij 2498242845Sdelphij if (zfs_disable_dup_eviction) 2499242845Sdelphij return (B_FALSE); 2500242845Sdelphij 2501242845Sdelphij mutex_enter(&buf->b_evict_lock); 2502242845Sdelphij hdr = buf->b_hdr; 2503242845Sdelphij if (hdr == NULL) { 2504242845Sdelphij /* 2505242845Sdelphij * We are in arc_do_user_evicts(); let that function 2506242845Sdelphij * perform the eviction. 2507242845Sdelphij */ 2508242845Sdelphij ASSERT(buf->b_data == NULL); 2509242845Sdelphij mutex_exit(&buf->b_evict_lock); 2510242845Sdelphij return (B_FALSE); 2511242845Sdelphij } else if (buf->b_data == NULL) { 2512242845Sdelphij /* 2513242845Sdelphij * We have already been added to the arc eviction list; 2514242845Sdelphij * recommend eviction. 2515242845Sdelphij */ 2516242845Sdelphij ASSERT3P(hdr, ==, &arc_eviction_hdr); 2517242845Sdelphij mutex_exit(&buf->b_evict_lock); 2518242845Sdelphij return (B_TRUE); 2519242845Sdelphij } 2520242845Sdelphij 2521286570Smav if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2522242845Sdelphij evict_needed = B_TRUE; 2523242845Sdelphij 2524242845Sdelphij mutex_exit(&buf->b_evict_lock); 2525242845Sdelphij return (evict_needed); 2526242845Sdelphij} 2527242845Sdelphij 2528242845Sdelphij/* 2529286763Smav * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2530286763Smav * state of the header is dependent on it's state prior to entering this 2531286763Smav * function. The following transitions are possible: 2532185029Spjd * 2533286763Smav * - arc_mru -> arc_mru_ghost 2534286763Smav * - arc_mfu -> arc_mfu_ghost 2535286763Smav * - arc_mru_ghost -> arc_l2c_only 2536286763Smav * - arc_mru_ghost -> deleted 2537286763Smav * - arc_mfu_ghost -> arc_l2c_only 2538286763Smav * - arc_mfu_ghost -> deleted 2539168404Spjd */ 2540286763Smavstatic int64_t 2541286763Smavarc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2542168404Spjd{ 2543286763Smav arc_state_t *evicted_state, *state; 2544286763Smav int64_t bytes_evicted = 0; 2545168404Spjd 2546286763Smav ASSERT(MUTEX_HELD(hash_lock)); 2547286763Smav ASSERT(HDR_HAS_L1HDR(hdr)); 2548168404Spjd 2549286763Smav state = hdr->b_l1hdr.b_state; 2550286763Smav if (GHOST_STATE(state)) { 2551286763Smav ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2552286763Smav ASSERT(hdr->b_l1hdr.b_buf == NULL); 2553206796Spjd 2554286763Smav /* 2555286763Smav * l2arc_write_buffers() relies on a header's L1 portion 2556286763Smav * (i.e. it's b_tmp_cdata field) during it's write phase. 2557286763Smav * Thus, we cannot push a header onto the arc_l2c_only 2558286763Smav * state (removing it's L1 piece) until the header is 2559286763Smav * done being written to the l2arc. 2560286763Smav */ 2561286763Smav if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2562286763Smav ARCSTAT_BUMP(arcstat_evict_l2_skip); 2563286763Smav return (bytes_evicted); 2564286763Smav } 2565286762Smav 2566286763Smav ARCSTAT_BUMP(arcstat_deleted); 2567286763Smav bytes_evicted += hdr->b_size; 2568286762Smav 2569286763Smav DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2570286763Smav 2571286763Smav if (HDR_HAS_L2HDR(hdr)) { 2572275780Sdelphij /* 2573286763Smav * This buffer is cached on the 2nd Level ARC; 2574286763Smav * don't destroy the header. 2575275780Sdelphij */ 2576286763Smav arc_change_state(arc_l2c_only, hdr, hash_lock); 2577286763Smav /* 2578286763Smav * dropping from L1+L2 cached to L2-only, 2579286763Smav * realloc to remove the L1 header. 2580286763Smav */ 2581286763Smav hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2582286763Smav hdr_l2only_cache); 2583286763Smav } else { 2584286763Smav arc_change_state(arc_anon, hdr, hash_lock); 2585286763Smav arc_hdr_destroy(hdr); 2586275780Sdelphij } 2587286763Smav return (bytes_evicted); 2588275780Sdelphij } 2589275780Sdelphij 2590286763Smav ASSERT(state == arc_mru || state == arc_mfu); 2591286763Smav evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2592206796Spjd 2593286763Smav /* prefetch buffers have a minimum lifespan */ 2594286763Smav if (HDR_IO_IN_PROGRESS(hdr) || 2595286763Smav ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2596286763Smav ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2597286763Smav arc_min_prefetch_lifespan)) { 2598286763Smav ARCSTAT_BUMP(arcstat_evict_skip); 2599286763Smav return (bytes_evicted); 2600286763Smav } 2601286763Smav 2602286763Smav ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2603286763Smav ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2604286763Smav while (hdr->b_l1hdr.b_buf) { 2605286763Smav arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2606286763Smav if (!mutex_tryenter(&buf->b_evict_lock)) { 2607286763Smav ARCSTAT_BUMP(arcstat_mutex_miss); 2608286763Smav break; 2609168404Spjd } 2610286763Smav if (buf->b_data != NULL) 2611286763Smav bytes_evicted += hdr->b_size; 2612286763Smav if (buf->b_efunc != NULL) { 2613286763Smav mutex_enter(&arc_user_evicts_lock); 2614286763Smav arc_buf_destroy(buf, FALSE); 2615286763Smav hdr->b_l1hdr.b_buf = buf->b_next; 2616286763Smav buf->b_hdr = &arc_eviction_hdr; 2617286763Smav buf->b_next = arc_eviction_list; 2618286763Smav arc_eviction_list = buf; 2619286763Smav cv_signal(&arc_user_evicts_cv); 2620286763Smav mutex_exit(&arc_user_evicts_lock); 2621286763Smav mutex_exit(&buf->b_evict_lock); 2622286763Smav } else { 2623286763Smav mutex_exit(&buf->b_evict_lock); 2624286763Smav arc_buf_destroy(buf, TRUE); 2625286763Smav } 2626286763Smav } 2627258632Savg 2628286763Smav if (HDR_HAS_L2HDR(hdr)) { 2629286763Smav ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2630286763Smav } else { 2631286763Smav if (l2arc_write_eligible(hdr->b_spa, hdr)) 2632286763Smav ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2633286763Smav else 2634286763Smav ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2635286763Smav } 2636258632Savg 2637286763Smav if (hdr->b_l1hdr.b_datacnt == 0) { 2638286763Smav arc_change_state(evicted_state, hdr, hash_lock); 2639286763Smav ASSERT(HDR_IN_HASH_TABLE(hdr)); 2640286763Smav hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2641286763Smav hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2642286763Smav DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2643286763Smav } 2644286763Smav 2645286763Smav return (bytes_evicted); 2646286763Smav} 2647286763Smav 2648286763Smavstatic uint64_t 2649286763Smavarc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2650286763Smav uint64_t spa, int64_t bytes) 2651286763Smav{ 2652286763Smav multilist_sublist_t *mls; 2653286763Smav uint64_t bytes_evicted = 0; 2654286763Smav arc_buf_hdr_t *hdr; 2655286763Smav kmutex_t *hash_lock; 2656286763Smav int evict_count = 0; 2657286763Smav 2658286763Smav ASSERT3P(marker, !=, NULL); 2659286763Smav IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2660286763Smav 2661286763Smav mls = multilist_sublist_lock(ml, idx); 2662286763Smav 2663286763Smav for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2664286763Smav hdr = multilist_sublist_prev(mls, marker)) { 2665286763Smav if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2666286763Smav (evict_count >= zfs_arc_evict_batch_limit)) 2667286763Smav break; 2668286763Smav 2669258632Savg /* 2670286763Smav * To keep our iteration location, move the marker 2671286763Smav * forward. Since we're not holding hdr's hash lock, we 2672286763Smav * must be very careful and not remove 'hdr' from the 2673286763Smav * sublist. Otherwise, other consumers might mistake the 2674286763Smav * 'hdr' as not being on a sublist when they call the 2675286763Smav * multilist_link_active() function (they all rely on 2676286763Smav * the hash lock protecting concurrent insertions and 2677286763Smav * removals). multilist_sublist_move_forward() was 2678286763Smav * specifically implemented to ensure this is the case 2679286763Smav * (only 'marker' will be removed and re-inserted). 2680258632Savg */ 2681286763Smav multilist_sublist_move_forward(mls, marker); 2682286763Smav 2683286763Smav /* 2684286763Smav * The only case where the b_spa field should ever be 2685286763Smav * zero, is the marker headers inserted by 2686286763Smav * arc_evict_state(). It's possible for multiple threads 2687286763Smav * to be calling arc_evict_state() concurrently (e.g. 2688286763Smav * dsl_pool_close() and zio_inject_fault()), so we must 2689286763Smav * skip any markers we see from these other threads. 2690286763Smav */ 2691286763Smav if (hdr->b_spa == 0) 2692258632Savg continue; 2693286763Smav 2694286763Smav /* we're only interested in evicting buffers of a certain spa */ 2695286763Smav if (spa != 0 && hdr->b_spa != spa) { 2696286763Smav ARCSTAT_BUMP(arcstat_evict_skip); 2697286763Smav continue; 2698258632Savg } 2699258632Savg 2700275811Sdelphij hash_lock = HDR_LOCK(hdr); 2701208373Smm 2702286763Smav /* 2703286763Smav * We aren't calling this function from any code path 2704286763Smav * that would already be holding a hash lock, so we're 2705286763Smav * asserting on this assumption to be defensive in case 2706286763Smav * this ever changes. Without this check, it would be 2707286763Smav * possible to incorrectly increment arcstat_mutex_miss 2708286763Smav * below (e.g. if the code changed such that we called 2709286763Smav * this function with a hash lock held). 2710286763Smav */ 2711286763Smav ASSERT(!MUTEX_HELD(hash_lock)); 2712208373Smm 2713286763Smav if (mutex_tryenter(hash_lock)) { 2714286763Smav uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2715286763Smav mutex_exit(hash_lock); 2716286763Smav 2717286763Smav bytes_evicted += evicted; 2718286763Smav 2719286763Smav /* 2720286763Smav * If evicted is zero, arc_evict_hdr() must have 2721286763Smav * decided to skip this header, don't increment 2722286763Smav * evict_count in this case. 2723286763Smav */ 2724286763Smav if (evicted != 0) 2725286763Smav evict_count++; 2726286763Smav 2727286763Smav /* 2728286763Smav * If arc_size isn't overflowing, signal any 2729286763Smav * threads that might happen to be waiting. 2730286763Smav * 2731286763Smav * For each header evicted, we wake up a single 2732286763Smav * thread. If we used cv_broadcast, we could 2733286763Smav * wake up "too many" threads causing arc_size 2734286763Smav * to significantly overflow arc_c; since 2735286763Smav * arc_get_data_buf() doesn't check for overflow 2736286763Smav * when it's woken up (it doesn't because it's 2737286763Smav * possible for the ARC to be overflowing while 2738286763Smav * full of un-evictable buffers, and the 2739286763Smav * function should proceed in this case). 2740286763Smav * 2741286763Smav * If threads are left sleeping, due to not 2742286763Smav * using cv_broadcast, they will be woken up 2743286763Smav * just before arc_reclaim_thread() sleeps. 2744286763Smav */ 2745286763Smav mutex_enter(&arc_reclaim_lock); 2746286763Smav if (!arc_is_overflowing()) 2747286763Smav cv_signal(&arc_reclaim_waiters_cv); 2748286763Smav mutex_exit(&arc_reclaim_lock); 2749168404Spjd } else { 2750286763Smav ARCSTAT_BUMP(arcstat_mutex_miss); 2751168404Spjd } 2752168404Spjd } 2753168404Spjd 2754286763Smav multilist_sublist_unlock(mls); 2755206796Spjd 2756286763Smav return (bytes_evicted); 2757286763Smav} 2758168404Spjd 2759286763Smav/* 2760286763Smav * Evict buffers from the given arc state, until we've removed the 2761286763Smav * specified number of bytes. Move the removed buffers to the 2762286763Smav * appropriate evict state. 2763286763Smav * 2764286763Smav * This function makes a "best effort". It skips over any buffers 2765286763Smav * it can't get a hash_lock on, and so, may not catch all candidates. 2766286763Smav * It may also return without evicting as much space as requested. 2767286763Smav * 2768286763Smav * If bytes is specified using the special value ARC_EVICT_ALL, this 2769286763Smav * will evict all available (i.e. unlocked and evictable) buffers from 2770286763Smav * the given arc state; which is used by arc_flush(). 2771286763Smav */ 2772286763Smavstatic uint64_t 2773286763Smavarc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2774286763Smav arc_buf_contents_t type) 2775286763Smav{ 2776286763Smav uint64_t total_evicted = 0; 2777286763Smav multilist_t *ml = &state->arcs_list[type]; 2778286763Smav int num_sublists; 2779286763Smav arc_buf_hdr_t **markers; 2780168404Spjd 2781286763Smav IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2782168404Spjd 2783286763Smav num_sublists = multilist_get_num_sublists(ml); 2784286763Smav 2785185029Spjd /* 2786286763Smav * If we've tried to evict from each sublist, made some 2787286763Smav * progress, but still have not hit the target number of bytes 2788286763Smav * to evict, we want to keep trying. The markers allow us to 2789286763Smav * pick up where we left off for each individual sublist, rather 2790286763Smav * than starting from the tail each time. 2791185029Spjd */ 2792286763Smav markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2793286763Smav for (int i = 0; i < num_sublists; i++) { 2794286763Smav markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2795185029Spjd 2796286763Smav /* 2797286763Smav * A b_spa of 0 is used to indicate that this header is 2798286763Smav * a marker. This fact is used in arc_adjust_type() and 2799286763Smav * arc_evict_state_impl(). 2800286763Smav */ 2801286763Smav markers[i]->b_spa = 0; 2802168404Spjd 2803286763Smav multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2804286763Smav multilist_sublist_insert_tail(mls, markers[i]); 2805286763Smav multilist_sublist_unlock(mls); 2806286763Smav } 2807168404Spjd 2808286763Smav /* 2809286763Smav * While we haven't hit our target number of bytes to evict, or 2810286763Smav * we're evicting all available buffers. 2811286763Smav */ 2812286763Smav while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2813286763Smav /* 2814286763Smav * Start eviction using a randomly selected sublist, 2815286763Smav * this is to try and evenly balance eviction across all 2816286763Smav * sublists. Always starting at the same sublist 2817286763Smav * (e.g. index 0) would cause evictions to favor certain 2818286763Smav * sublists over others. 2819286763Smav */ 2820286763Smav int sublist_idx = multilist_get_random_index(ml); 2821286763Smav uint64_t scan_evicted = 0; 2822219089Spjd 2823286763Smav for (int i = 0; i < num_sublists; i++) { 2824286763Smav uint64_t bytes_remaining; 2825286763Smav uint64_t bytes_evicted; 2826219089Spjd 2827286763Smav if (bytes == ARC_EVICT_ALL) 2828286763Smav bytes_remaining = ARC_EVICT_ALL; 2829286763Smav else if (total_evicted < bytes) 2830286763Smav bytes_remaining = bytes - total_evicted; 2831286763Smav else 2832286763Smav break; 2833258632Savg 2834286763Smav bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2835286763Smav markers[sublist_idx], spa, bytes_remaining); 2836286763Smav 2837286763Smav scan_evicted += bytes_evicted; 2838286763Smav total_evicted += bytes_evicted; 2839286763Smav 2840286763Smav /* we've reached the end, wrap to the beginning */ 2841286763Smav if (++sublist_idx >= num_sublists) 2842286763Smav sublist_idx = 0; 2843286763Smav } 2844286763Smav 2845258632Savg /* 2846286763Smav * If we didn't evict anything during this scan, we have 2847286763Smav * no reason to believe we'll evict more during another 2848286763Smav * scan, so break the loop. 2849258632Savg */ 2850286763Smav if (scan_evicted == 0) { 2851286763Smav /* This isn't possible, let's make that obvious */ 2852286763Smav ASSERT3S(bytes, !=, 0); 2853185029Spjd 2854286763Smav /* 2855286763Smav * When bytes is ARC_EVICT_ALL, the only way to 2856286763Smav * break the loop is when scan_evicted is zero. 2857286763Smav * In that case, we actually have evicted enough, 2858286763Smav * so we don't want to increment the kstat. 2859286763Smav */ 2860286763Smav if (bytes != ARC_EVICT_ALL) { 2861286763Smav ASSERT3S(total_evicted, <, bytes); 2862286763Smav ARCSTAT_BUMP(arcstat_evict_not_enough); 2863185029Spjd } 2864185029Spjd 2865286763Smav break; 2866258632Savg } 2867286763Smav } 2868258632Savg 2869286763Smav for (int i = 0; i < num_sublists; i++) { 2870286763Smav multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2871286763Smav multilist_sublist_remove(mls, markers[i]); 2872286763Smav multilist_sublist_unlock(mls); 2873286763Smav 2874286763Smav kmem_cache_free(hdr_full_cache, markers[i]); 2875168404Spjd } 2876286763Smav kmem_free(markers, sizeof (*markers) * num_sublists); 2877206796Spjd 2878286763Smav return (total_evicted); 2879286763Smav} 2880286763Smav 2881286763Smav/* 2882286763Smav * Flush all "evictable" data of the given type from the arc state 2883286763Smav * specified. This will not evict any "active" buffers (i.e. referenced). 2884286763Smav * 2885286763Smav * When 'retry' is set to FALSE, the function will make a single pass 2886286763Smav * over the state and evict any buffers that it can. Since it doesn't 2887286763Smav * continually retry the eviction, it might end up leaving some buffers 2888286763Smav * in the ARC due to lock misses. 2889286763Smav * 2890286763Smav * When 'retry' is set to TRUE, the function will continually retry the 2891286763Smav * eviction until *all* evictable buffers have been removed from the 2892286763Smav * state. As a result, if concurrent insertions into the state are 2893286763Smav * allowed (e.g. if the ARC isn't shutting down), this function might 2894286763Smav * wind up in an infinite loop, continually trying to evict buffers. 2895286763Smav */ 2896286763Smavstatic uint64_t 2897286763Smavarc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2898286763Smav boolean_t retry) 2899286763Smav{ 2900286763Smav uint64_t evicted = 0; 2901286763Smav 2902286763Smav while (state->arcs_lsize[type] != 0) { 2903286763Smav evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2904286763Smav 2905286763Smav if (!retry) 2906286763Smav break; 2907185029Spjd } 2908185029Spjd 2909286763Smav return (evicted); 2910286763Smav} 2911286763Smav 2912286763Smav/* 2913286763Smav * Evict the specified number of bytes from the state specified, 2914286763Smav * restricting eviction to the spa and type given. This function 2915286763Smav * prevents us from trying to evict more from a state's list than 2916286763Smav * is "evictable", and to skip evicting altogether when passed a 2917286763Smav * negative value for "bytes". In contrast, arc_evict_state() will 2918286763Smav * evict everything it can, when passed a negative value for "bytes". 2919286763Smav */ 2920286763Smavstatic uint64_t 2921286763Smavarc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 2922286763Smav arc_buf_contents_t type) 2923286763Smav{ 2924286763Smav int64_t delta; 2925286763Smav 2926286763Smav if (bytes > 0 && state->arcs_lsize[type] > 0) { 2927286763Smav delta = MIN(state->arcs_lsize[type], bytes); 2928286763Smav return (arc_evict_state(state, spa, delta, type)); 2929168404Spjd } 2930168404Spjd 2931286763Smav return (0); 2932168404Spjd} 2933168404Spjd 2934286763Smav/* 2935286763Smav * Evict metadata buffers from the cache, such that arc_meta_used is 2936286763Smav * capped by the arc_meta_limit tunable. 2937286763Smav */ 2938286763Smavstatic uint64_t 2939286763Smavarc_adjust_meta(void) 2940286763Smav{ 2941286763Smav uint64_t total_evicted = 0; 2942286763Smav int64_t target; 2943286763Smav 2944286763Smav /* 2945286763Smav * If we're over the meta limit, we want to evict enough 2946286763Smav * metadata to get back under the meta limit. We don't want to 2947286763Smav * evict so much that we drop the MRU below arc_p, though. If 2948286763Smav * we're over the meta limit more than we're over arc_p, we 2949286763Smav * evict some from the MRU here, and some from the MFU below. 2950286763Smav */ 2951286763Smav target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 2952286763Smav (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size - arc_p)); 2953286763Smav 2954286763Smav total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 2955286763Smav 2956286763Smav /* 2957286763Smav * Similar to the above, we want to evict enough bytes to get us 2958286763Smav * below the meta limit, but not so much as to drop us below the 2959286763Smav * space alloted to the MFU (which is defined as arc_c - arc_p). 2960286763Smav */ 2961286763Smav target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 2962286763Smav (int64_t)(arc_mfu->arcs_size - (arc_c - arc_p))); 2963286763Smav 2964286763Smav total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 2965286763Smav 2966286763Smav return (total_evicted); 2967286763Smav} 2968286763Smav 2969286763Smav/* 2970286763Smav * Return the type of the oldest buffer in the given arc state 2971286763Smav * 2972286763Smav * This function will select a random sublist of type ARC_BUFC_DATA and 2973286763Smav * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 2974286763Smav * is compared, and the type which contains the "older" buffer will be 2975286763Smav * returned. 2976286763Smav */ 2977286763Smavstatic arc_buf_contents_t 2978286763Smavarc_adjust_type(arc_state_t *state) 2979286763Smav{ 2980286763Smav multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 2981286763Smav multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 2982286763Smav int data_idx = multilist_get_random_index(data_ml); 2983286763Smav int meta_idx = multilist_get_random_index(meta_ml); 2984286763Smav multilist_sublist_t *data_mls; 2985286763Smav multilist_sublist_t *meta_mls; 2986286763Smav arc_buf_contents_t type; 2987286763Smav arc_buf_hdr_t *data_hdr; 2988286763Smav arc_buf_hdr_t *meta_hdr; 2989286763Smav 2990286763Smav /* 2991286763Smav * We keep the sublist lock until we're finished, to prevent 2992286763Smav * the headers from being destroyed via arc_evict_state(). 2993286763Smav */ 2994286763Smav data_mls = multilist_sublist_lock(data_ml, data_idx); 2995286763Smav meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 2996286763Smav 2997286763Smav /* 2998286763Smav * These two loops are to ensure we skip any markers that 2999286763Smav * might be at the tail of the lists due to arc_evict_state(). 3000286763Smav */ 3001286763Smav 3002286763Smav for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3003286763Smav data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3004286763Smav if (data_hdr->b_spa != 0) 3005286763Smav break; 3006286763Smav } 3007286763Smav 3008286763Smav for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3009286763Smav meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3010286763Smav if (meta_hdr->b_spa != 0) 3011286763Smav break; 3012286763Smav } 3013286763Smav 3014286763Smav if (data_hdr == NULL && meta_hdr == NULL) { 3015286763Smav type = ARC_BUFC_DATA; 3016286763Smav } else if (data_hdr == NULL) { 3017286763Smav ASSERT3P(meta_hdr, !=, NULL); 3018286763Smav type = ARC_BUFC_METADATA; 3019286763Smav } else if (meta_hdr == NULL) { 3020286763Smav ASSERT3P(data_hdr, !=, NULL); 3021286763Smav type = ARC_BUFC_DATA; 3022286763Smav } else { 3023286763Smav ASSERT3P(data_hdr, !=, NULL); 3024286763Smav ASSERT3P(meta_hdr, !=, NULL); 3025286763Smav 3026286763Smav /* The headers can't be on the sublist without an L1 header */ 3027286763Smav ASSERT(HDR_HAS_L1HDR(data_hdr)); 3028286763Smav ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3029286763Smav 3030286763Smav if (data_hdr->b_l1hdr.b_arc_access < 3031286763Smav meta_hdr->b_l1hdr.b_arc_access) { 3032286763Smav type = ARC_BUFC_DATA; 3033286763Smav } else { 3034286763Smav type = ARC_BUFC_METADATA; 3035286763Smav } 3036286763Smav } 3037286763Smav 3038286763Smav multilist_sublist_unlock(meta_mls); 3039286763Smav multilist_sublist_unlock(data_mls); 3040286763Smav 3041286763Smav return (type); 3042286763Smav} 3043286763Smav 3044286763Smav/* 3045286763Smav * Evict buffers from the cache, such that arc_size is capped by arc_c. 3046286763Smav */ 3047286763Smavstatic uint64_t 3048168404Spjdarc_adjust(void) 3049168404Spjd{ 3050286763Smav uint64_t total_evicted = 0; 3051286763Smav uint64_t bytes; 3052286763Smav int64_t target; 3053168404Spjd 3054208373Smm /* 3055286763Smav * If we're over arc_meta_limit, we want to correct that before 3056286763Smav * potentially evicting data buffers below. 3057286763Smav */ 3058286763Smav total_evicted += arc_adjust_meta(); 3059286763Smav 3060286763Smav /* 3061208373Smm * Adjust MRU size 3062286763Smav * 3063286763Smav * If we're over the target cache size, we want to evict enough 3064286763Smav * from the list to get back to our target size. We don't want 3065286763Smav * to evict too much from the MRU, such that it drops below 3066286763Smav * arc_p. So, if we're over our target cache size more than 3067286763Smav * the MRU is over arc_p, we'll evict enough to get back to 3068286763Smav * arc_p here, and then evict more from the MFU below. 3069208373Smm */ 3070286763Smav target = MIN((int64_t)(arc_size - arc_c), 3071209275Smm (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 3072209275Smm arc_p)); 3073208373Smm 3074286763Smav /* 3075286763Smav * If we're below arc_meta_min, always prefer to evict data. 3076286763Smav * Otherwise, try to satisfy the requested number of bytes to 3077286763Smav * evict from the type which contains older buffers; in an 3078286763Smav * effort to keep newer buffers in the cache regardless of their 3079286763Smav * type. If we cannot satisfy the number of bytes from this 3080286763Smav * type, spill over into the next type. 3081286763Smav */ 3082286763Smav if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3083286763Smav arc_meta_used > arc_meta_min) { 3084286763Smav bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3085286763Smav total_evicted += bytes; 3086168404Spjd 3087286763Smav /* 3088286763Smav * If we couldn't evict our target number of bytes from 3089286763Smav * metadata, we try to get the rest from data. 3090286763Smav */ 3091286763Smav target -= bytes; 3092286763Smav 3093286763Smav total_evicted += 3094286763Smav arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3095286763Smav } else { 3096286763Smav bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3097286763Smav total_evicted += bytes; 3098286763Smav 3099286763Smav /* 3100286763Smav * If we couldn't evict our target number of bytes from 3101286763Smav * data, we try to get the rest from metadata. 3102286763Smav */ 3103286763Smav target -= bytes; 3104286763Smav 3105286763Smav total_evicted += 3106286763Smav arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3107185029Spjd } 3108185029Spjd 3109208373Smm /* 3110208373Smm * Adjust MFU size 3111286763Smav * 3112286763Smav * Now that we've tried to evict enough from the MRU to get its 3113286763Smav * size back to arc_p, if we're still above the target cache 3114286763Smav * size, we evict the rest from the MFU. 3115208373Smm */ 3116286763Smav target = arc_size - arc_c; 3117168404Spjd 3118286764Smav if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3119286763Smav arc_meta_used > arc_meta_min) { 3120286763Smav bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3121286763Smav total_evicted += bytes; 3122208373Smm 3123286763Smav /* 3124286763Smav * If we couldn't evict our target number of bytes from 3125286763Smav * metadata, we try to get the rest from data. 3126286763Smav */ 3127286763Smav target -= bytes; 3128168404Spjd 3129286763Smav total_evicted += 3130286763Smav arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3131286763Smav } else { 3132286763Smav bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3133286763Smav total_evicted += bytes; 3134286763Smav 3135286763Smav /* 3136286763Smav * If we couldn't evict our target number of bytes from 3137286763Smav * data, we try to get the rest from data. 3138286763Smav */ 3139286763Smav target -= bytes; 3140286763Smav 3141286763Smav total_evicted += 3142286763Smav arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3143208373Smm } 3144168404Spjd 3145208373Smm /* 3146208373Smm * Adjust ghost lists 3147286763Smav * 3148286763Smav * In addition to the above, the ARC also defines target values 3149286763Smav * for the ghost lists. The sum of the mru list and mru ghost 3150286763Smav * list should never exceed the target size of the cache, and 3151286763Smav * the sum of the mru list, mfu list, mru ghost list, and mfu 3152286763Smav * ghost list should never exceed twice the target size of the 3153286763Smav * cache. The following logic enforces these limits on the ghost 3154286763Smav * caches, and evicts from them as needed. 3155208373Smm */ 3156286763Smav target = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 3157168404Spjd 3158286763Smav bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3159286763Smav total_evicted += bytes; 3160168404Spjd 3161286763Smav target -= bytes; 3162185029Spjd 3163286763Smav total_evicted += 3164286763Smav arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3165208373Smm 3166286763Smav /* 3167286763Smav * We assume the sum of the mru list and mfu list is less than 3168286763Smav * or equal to arc_c (we enforced this above), which means we 3169286763Smav * can use the simpler of the two equations below: 3170286763Smav * 3171286763Smav * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3172286763Smav * mru ghost + mfu ghost <= arc_c 3173286763Smav */ 3174286763Smav target = arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 3175286763Smav 3176286763Smav bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3177286763Smav total_evicted += bytes; 3178286763Smav 3179286763Smav target -= bytes; 3180286763Smav 3181286763Smav total_evicted += 3182286763Smav arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3183286763Smav 3184286763Smav return (total_evicted); 3185168404Spjd} 3186168404Spjd 3187168404Spjdstatic void 3188168404Spjdarc_do_user_evicts(void) 3189168404Spjd{ 3190286763Smav mutex_enter(&arc_user_evicts_lock); 3191286762Smav while (arc_eviction_list != NULL) { 3192286762Smav arc_buf_t *buf = arc_eviction_list; 3193286762Smav arc_eviction_list = buf->b_next; 3194219089Spjd mutex_enter(&buf->b_evict_lock); 3195168404Spjd buf->b_hdr = NULL; 3196219089Spjd mutex_exit(&buf->b_evict_lock); 3197286763Smav mutex_exit(&arc_user_evicts_lock); 3198168404Spjd 3199168404Spjd if (buf->b_efunc != NULL) 3200268858Sdelphij VERIFY0(buf->b_efunc(buf->b_private)); 3201168404Spjd 3202168404Spjd buf->b_efunc = NULL; 3203168404Spjd buf->b_private = NULL; 3204168404Spjd kmem_cache_free(buf_cache, buf); 3205286763Smav mutex_enter(&arc_user_evicts_lock); 3206168404Spjd } 3207286763Smav mutex_exit(&arc_user_evicts_lock); 3208168404Spjd} 3209168404Spjd 3210168404Spjdvoid 3211286763Smavarc_flush(spa_t *spa, boolean_t retry) 3212168404Spjd{ 3213209962Smm uint64_t guid = 0; 3214209962Smm 3215286763Smav /* 3216286763Smav * If retry is TRUE, a spa must not be specified since we have 3217286763Smav * no good way to determine if all of a spa's buffers have been 3218286763Smav * evicted from an arc state. 3219286763Smav */ 3220286763Smav ASSERT(!retry || spa == 0); 3221286763Smav 3222286570Smav if (spa != NULL) 3223228103Smm guid = spa_load_guid(spa); 3224209962Smm 3225286763Smav (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3226286763Smav (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3227168404Spjd 3228286763Smav (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3229286763Smav (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3230168404Spjd 3231286763Smav (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3232286763Smav (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3233286763Smav 3234286763Smav (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3235286763Smav (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3236286763Smav 3237168404Spjd arc_do_user_evicts(); 3238185029Spjd ASSERT(spa || arc_eviction_list == NULL); 3239168404Spjd} 3240168404Spjd 3241168404Spjdvoid 3242286625Smavarc_shrink(int64_t to_free) 3243168404Spjd{ 3244168404Spjd if (arc_c > arc_c_min) { 3245272483Ssmh DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3246272483Ssmh arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3247168404Spjd if (arc_c > arc_c_min + to_free) 3248168404Spjd atomic_add_64(&arc_c, -to_free); 3249168404Spjd else 3250168404Spjd arc_c = arc_c_min; 3251168404Spjd 3252168404Spjd atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3253168404Spjd if (arc_c > arc_size) 3254168404Spjd arc_c = MAX(arc_size, arc_c_min); 3255168404Spjd if (arc_p > arc_c) 3256168404Spjd arc_p = (arc_c >> 1); 3257272483Ssmh 3258272483Ssmh DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3259272483Ssmh arc_p); 3260272483Ssmh 3261168404Spjd ASSERT(arc_c >= arc_c_min); 3262168404Spjd ASSERT((int64_t)arc_p >= 0); 3263168404Spjd } 3264168404Spjd 3265270759Ssmh if (arc_size > arc_c) { 3266270759Ssmh DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3267270759Ssmh uint64_t, arc_c); 3268286763Smav (void) arc_adjust(); 3269270759Ssmh } 3270168404Spjd} 3271168404Spjd 3272286625Smavstatic long needfree = 0; 3273168404Spjd 3274286625Smavtypedef enum free_memory_reason_t { 3275286625Smav FMR_UNKNOWN, 3276286625Smav FMR_NEEDFREE, 3277286625Smav FMR_LOTSFREE, 3278286625Smav FMR_SWAPFS_MINFREE, 3279286625Smav FMR_PAGES_PP_MAXIMUM, 3280286625Smav FMR_HEAP_ARENA, 3281286625Smav FMR_ZIO_ARENA, 3282286625Smav FMR_ZIO_FRAG, 3283286625Smav} free_memory_reason_t; 3284286625Smav 3285286625Smavint64_t last_free_memory; 3286286625Smavfree_memory_reason_t last_free_reason; 3287286625Smav 3288286625Smav/* 3289286625Smav * Additional reserve of pages for pp_reserve. 3290286625Smav */ 3291286625Smavint64_t arc_pages_pp_reserve = 64; 3292286625Smav 3293286625Smav/* 3294286625Smav * Additional reserve of pages for swapfs. 3295286625Smav */ 3296286625Smavint64_t arc_swapfs_reserve = 64; 3297286625Smav 3298286625Smav/* 3299286625Smav * Return the amount of memory that can be consumed before reclaim will be 3300286625Smav * needed. Positive if there is sufficient free memory, negative indicates 3301286625Smav * the amount of memory that needs to be freed up. 3302286625Smav */ 3303286625Smavstatic int64_t 3304286625Smavarc_available_memory(void) 3305168404Spjd{ 3306286625Smav int64_t lowest = INT64_MAX; 3307286625Smav int64_t n; 3308286625Smav free_memory_reason_t r = FMR_UNKNOWN; 3309168404Spjd 3310168404Spjd#ifdef _KERNEL 3311286625Smav if (needfree > 0) { 3312286625Smav n = PAGESIZE * (-needfree); 3313286625Smav if (n < lowest) { 3314286625Smav lowest = n; 3315286625Smav r = FMR_NEEDFREE; 3316286625Smav } 3317270759Ssmh } 3318168404Spjd 3319191902Skmacy /* 3320212780Savg * Cooperate with pagedaemon when it's time for it to scan 3321212780Savg * and reclaim some pages. 3322191902Skmacy */ 3323286655Smav n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3324286625Smav if (n < lowest) { 3325286625Smav lowest = n; 3326286625Smav r = FMR_LOTSFREE; 3327270759Ssmh } 3328191902Skmacy 3329277300Ssmh#ifdef illumos 3330168404Spjd /* 3331185029Spjd * check that we're out of range of the pageout scanner. It starts to 3332185029Spjd * schedule paging if freemem is less than lotsfree and needfree. 3333185029Spjd * lotsfree is the high-water mark for pageout, and needfree is the 3334185029Spjd * number of needed free pages. We add extra pages here to make sure 3335185029Spjd * the scanner doesn't start up while we're freeing memory. 3336185029Spjd */ 3337286625Smav n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3338286625Smav if (n < lowest) { 3339286625Smav lowest = n; 3340286625Smav r = FMR_LOTSFREE; 3341286625Smav } 3342185029Spjd 3343185029Spjd /* 3344168404Spjd * check to make sure that swapfs has enough space so that anon 3345185029Spjd * reservations can still succeed. anon_resvmem() checks that the 3346168404Spjd * availrmem is greater than swapfs_minfree, and the number of reserved 3347168404Spjd * swap pages. We also add a bit of extra here just to prevent 3348168404Spjd * circumstances from getting really dire. 3349168404Spjd */ 3350286625Smav n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3351286625Smav desfree - arc_swapfs_reserve); 3352286625Smav if (n < lowest) { 3353286625Smav lowest = n; 3354286625Smav r = FMR_SWAPFS_MINFREE; 3355286625Smav } 3356168404Spjd 3357286625Smav 3358168404Spjd /* 3359272483Ssmh * Check that we have enough availrmem that memory locking (e.g., via 3360272483Ssmh * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3361272483Ssmh * stores the number of pages that cannot be locked; when availrmem 3362272483Ssmh * drops below pages_pp_maximum, page locking mechanisms such as 3363272483Ssmh * page_pp_lock() will fail.) 3364272483Ssmh */ 3365286625Smav n = PAGESIZE * (availrmem - pages_pp_maximum - 3366286625Smav arc_pages_pp_reserve); 3367286625Smav if (n < lowest) { 3368286625Smav lowest = n; 3369286625Smav r = FMR_PAGES_PP_MAXIMUM; 3370286625Smav } 3371272483Ssmh 3372277300Ssmh#endif /* illumos */ 3373272483Ssmh#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3374272483Ssmh /* 3375168404Spjd * If we're on an i386 platform, it's possible that we'll exhaust the 3376168404Spjd * kernel heap space before we ever run out of available physical 3377168404Spjd * memory. Most checks of the size of the heap_area compare against 3378168404Spjd * tune.t_minarmem, which is the minimum available real memory that we 3379168404Spjd * can have in the system. However, this is generally fixed at 25 pages 3380168404Spjd * which is so low that it's useless. In this comparison, we seek to 3381168404Spjd * calculate the total heap-size, and reclaim if more than 3/4ths of the 3382185029Spjd * heap is allocated. (Or, in the calculation, if less than 1/4th is 3383168404Spjd * free) 3384168404Spjd */ 3385286655Smav n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3386286628Smav (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3387286625Smav if (n < lowest) { 3388286625Smav lowest = n; 3389286625Smav r = FMR_HEAP_ARENA; 3390270861Ssmh } 3391281026Smav#define zio_arena NULL 3392281026Smav#else 3393281026Smav#define zio_arena heap_arena 3394270861Ssmh#endif 3395281026Smav 3396272483Ssmh /* 3397272483Ssmh * If zio data pages are being allocated out of a separate heap segment, 3398272483Ssmh * then enforce that the size of available vmem for this arena remains 3399272483Ssmh * above about 1/16th free. 3400272483Ssmh * 3401272483Ssmh * Note: The 1/16th arena free requirement was put in place 3402272483Ssmh * to aggressively evict memory from the arc in order to avoid 3403272483Ssmh * memory fragmentation issues. 3404272483Ssmh */ 3405286625Smav if (zio_arena != NULL) { 3406286655Smav n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3407286625Smav (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3408286625Smav if (n < lowest) { 3409286625Smav lowest = n; 3410286625Smav r = FMR_ZIO_ARENA; 3411286625Smav } 3412286625Smav } 3413281026Smav 3414281026Smav /* 3415281026Smav * Above limits know nothing about real level of KVA fragmentation. 3416281026Smav * Start aggressive reclamation if too little sequential KVA left. 3417281026Smav */ 3418286625Smav if (lowest > 0) { 3419286625Smav n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3420286655Smav -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3421286655Smav INT64_MAX; 3422286625Smav if (n < lowest) { 3423286625Smav lowest = n; 3424286625Smav r = FMR_ZIO_FRAG; 3425286625Smav } 3426281109Smav } 3427281026Smav 3428272483Ssmh#else /* _KERNEL */ 3429286625Smav /* Every 100 calls, free a small amount */ 3430168404Spjd if (spa_get_random(100) == 0) 3431286625Smav lowest = -1024; 3432272483Ssmh#endif /* _KERNEL */ 3433270759Ssmh 3434286625Smav last_free_memory = lowest; 3435286625Smav last_free_reason = r; 3436286625Smav DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3437286625Smav return (lowest); 3438168404Spjd} 3439168404Spjd 3440286625Smav 3441286625Smav/* 3442286625Smav * Determine if the system is under memory pressure and is asking 3443286625Smav * to reclaim memory. A return value of TRUE indicates that the system 3444286625Smav * is under memory pressure and that the arc should adjust accordingly. 3445286625Smav */ 3446286625Smavstatic boolean_t 3447286625Smavarc_reclaim_needed(void) 3448286625Smav{ 3449286625Smav return (arc_available_memory() < 0); 3450286625Smav} 3451286625Smav 3452208454Spjdextern kmem_cache_t *zio_buf_cache[]; 3453208454Spjdextern kmem_cache_t *zio_data_buf_cache[]; 3454272527Sdelphijextern kmem_cache_t *range_seg_cache; 3455208454Spjd 3456278040Ssmhstatic __noinline void 3457286625Smavarc_kmem_reap_now(void) 3458168404Spjd{ 3459168404Spjd size_t i; 3460168404Spjd kmem_cache_t *prev_cache = NULL; 3461168404Spjd kmem_cache_t *prev_data_cache = NULL; 3462168404Spjd 3463272483Ssmh DTRACE_PROBE(arc__kmem_reap_start); 3464168404Spjd#ifdef _KERNEL 3465185029Spjd if (arc_meta_used >= arc_meta_limit) { 3466185029Spjd /* 3467185029Spjd * We are exceeding our meta-data cache limit. 3468185029Spjd * Purge some DNLC entries to release holds on meta-data. 3469185029Spjd */ 3470185029Spjd dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3471185029Spjd } 3472168404Spjd#if defined(__i386) 3473168404Spjd /* 3474168404Spjd * Reclaim unused memory from all kmem caches. 3475168404Spjd */ 3476168404Spjd kmem_reap(); 3477168404Spjd#endif 3478168404Spjd#endif 3479168404Spjd 3480168404Spjd for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3481168404Spjd if (zio_buf_cache[i] != prev_cache) { 3482168404Spjd prev_cache = zio_buf_cache[i]; 3483168404Spjd kmem_cache_reap_now(zio_buf_cache[i]); 3484168404Spjd } 3485168404Spjd if (zio_data_buf_cache[i] != prev_data_cache) { 3486168404Spjd prev_data_cache = zio_data_buf_cache[i]; 3487168404Spjd kmem_cache_reap_now(zio_data_buf_cache[i]); 3488168404Spjd } 3489168404Spjd } 3490168404Spjd kmem_cache_reap_now(buf_cache); 3491286570Smav kmem_cache_reap_now(hdr_full_cache); 3492286570Smav kmem_cache_reap_now(hdr_l2only_cache); 3493272506Sdelphij kmem_cache_reap_now(range_seg_cache); 3494272483Ssmh 3495277300Ssmh#ifdef illumos 3496286625Smav if (zio_arena != NULL) { 3497286625Smav /* 3498286625Smav * Ask the vmem arena to reclaim unused memory from its 3499286625Smav * quantum caches. 3500286625Smav */ 3501272483Ssmh vmem_qcache_reap(zio_arena); 3502286625Smav } 3503272483Ssmh#endif 3504272483Ssmh DTRACE_PROBE(arc__kmem_reap_end); 3505168404Spjd} 3506168404Spjd 3507286763Smav/* 3508286763Smav * Threads can block in arc_get_data_buf() waiting for this thread to evict 3509286763Smav * enough data and signal them to proceed. When this happens, the threads in 3510286763Smav * arc_get_data_buf() are sleeping while holding the hash lock for their 3511286763Smav * particular arc header. Thus, we must be careful to never sleep on a 3512286763Smav * hash lock in this thread. This is to prevent the following deadlock: 3513286763Smav * 3514286763Smav * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3515286763Smav * waiting for the reclaim thread to signal it. 3516286763Smav * 3517286763Smav * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3518286763Smav * fails, and goes to sleep forever. 3519286763Smav * 3520286763Smav * This possible deadlock is avoided by always acquiring a hash lock 3521286763Smav * using mutex_tryenter() from arc_reclaim_thread(). 3522286763Smav */ 3523168404Spjdstatic void 3524168404Spjdarc_reclaim_thread(void *dummy __unused) 3525168404Spjd{ 3526168404Spjd clock_t growtime = 0; 3527168404Spjd callb_cpr_t cpr; 3528168404Spjd 3529286763Smav CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3530168404Spjd 3531286763Smav mutex_enter(&arc_reclaim_lock); 3532286763Smav while (!arc_reclaim_thread_exit) { 3533286625Smav int64_t free_memory = arc_available_memory(); 3534286763Smav uint64_t evicted = 0; 3535286763Smav 3536286763Smav mutex_exit(&arc_reclaim_lock); 3537286763Smav 3538286625Smav if (free_memory < 0) { 3539168404Spjd 3540286625Smav arc_no_grow = B_TRUE; 3541286625Smav arc_warm = B_TRUE; 3542168404Spjd 3543286625Smav /* 3544286625Smav * Wait at least zfs_grow_retry (default 60) seconds 3545286625Smav * before considering growing. 3546286625Smav */ 3547219089Spjd growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3548168404Spjd 3549286625Smav arc_kmem_reap_now(); 3550286625Smav 3551286625Smav /* 3552286625Smav * If we are still low on memory, shrink the ARC 3553286625Smav * so that we have arc_shrink_min free space. 3554286625Smav */ 3555286625Smav free_memory = arc_available_memory(); 3556286625Smav 3557286625Smav int64_t to_free = 3558286625Smav (arc_c >> arc_shrink_shift) - free_memory; 3559286625Smav if (to_free > 0) { 3560286625Smav#ifdef _KERNEL 3561286625Smav to_free = MAX(to_free, ptob(needfree)); 3562286625Smav#endif 3563286625Smav arc_shrink(to_free); 3564168404Spjd } 3565286625Smav } else if (free_memory < arc_c >> arc_no_grow_shift) { 3566286625Smav arc_no_grow = B_TRUE; 3567286625Smav } else if (ddi_get_lbolt() >= growtime) { 3568286625Smav arc_no_grow = B_FALSE; 3569168404Spjd } 3570168404Spjd 3571286763Smav evicted = arc_adjust(); 3572168404Spjd 3573286763Smav mutex_enter(&arc_reclaim_lock); 3574168404Spjd 3575286763Smav /* 3576286763Smav * If evicted is zero, we couldn't evict anything via 3577286763Smav * arc_adjust(). This could be due to hash lock 3578286763Smav * collisions, but more likely due to the majority of 3579286763Smav * arc buffers being unevictable. Therefore, even if 3580286763Smav * arc_size is above arc_c, another pass is unlikely to 3581286763Smav * be helpful and could potentially cause us to enter an 3582286763Smav * infinite loop. 3583286763Smav */ 3584286763Smav if (arc_size <= arc_c || evicted == 0) { 3585211762Savg#ifdef _KERNEL 3586185029Spjd needfree = 0; 3587168404Spjd#endif 3588286763Smav /* 3589286763Smav * We're either no longer overflowing, or we 3590286763Smav * can't evict anything more, so we should wake 3591286763Smav * up any threads before we go to sleep. 3592286763Smav */ 3593286763Smav cv_broadcast(&arc_reclaim_waiters_cv); 3594168404Spjd 3595286763Smav /* 3596286763Smav * Block until signaled, or after one second (we 3597286763Smav * might need to perform arc_kmem_reap_now() 3598286763Smav * even if we aren't being signalled) 3599286763Smav */ 3600286763Smav CALLB_CPR_SAFE_BEGIN(&cpr); 3601286763Smav (void) cv_timedwait(&arc_reclaim_thread_cv, 3602286763Smav &arc_reclaim_lock, hz); 3603286763Smav CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3604286763Smav } 3605286763Smav } 3606286763Smav 3607286763Smav arc_reclaim_thread_exit = FALSE; 3608286763Smav cv_broadcast(&arc_reclaim_thread_cv); 3609286763Smav CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3610286763Smav thread_exit(); 3611286763Smav} 3612286763Smav 3613286763Smavstatic void 3614286763Smavarc_user_evicts_thread(void *dummy __unused) 3615286763Smav{ 3616286763Smav callb_cpr_t cpr; 3617286763Smav 3618286763Smav CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3619286763Smav 3620286763Smav mutex_enter(&arc_user_evicts_lock); 3621286763Smav while (!arc_user_evicts_thread_exit) { 3622286763Smav mutex_exit(&arc_user_evicts_lock); 3623286763Smav 3624286763Smav arc_do_user_evicts(); 3625286763Smav 3626286574Smav /* 3627286574Smav * This is necessary in order for the mdb ::arc dcmd to 3628286574Smav * show up to date information. Since the ::arc command 3629286574Smav * does not call the kstat's update function, without 3630286574Smav * this call, the command may show stale stats for the 3631286574Smav * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3632286574Smav * with this change, the data might be up to 1 second 3633286574Smav * out of date; but that should suffice. The arc_state_t 3634286574Smav * structures can be queried directly if more accurate 3635286574Smav * information is needed. 3636286574Smav */ 3637286574Smav if (arc_ksp != NULL) 3638286574Smav arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3639286574Smav 3640286763Smav mutex_enter(&arc_user_evicts_lock); 3641286763Smav 3642286763Smav /* 3643286763Smav * Block until signaled, or after one second (we need to 3644286763Smav * call the arc's kstat update function regularly). 3645286763Smav */ 3646168404Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 3647286763Smav (void) cv_timedwait(&arc_user_evicts_cv, 3648286763Smav &arc_user_evicts_lock, hz); 3649286763Smav CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3650168404Spjd } 3651168404Spjd 3652286763Smav arc_user_evicts_thread_exit = FALSE; 3653286763Smav cv_broadcast(&arc_user_evicts_cv); 3654286763Smav CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3655168404Spjd thread_exit(); 3656168404Spjd} 3657168404Spjd 3658168404Spjd/* 3659168404Spjd * Adapt arc info given the number of bytes we are trying to add and 3660168404Spjd * the state that we are comming from. This function is only called 3661168404Spjd * when we are adding new content to the cache. 3662168404Spjd */ 3663168404Spjdstatic void 3664168404Spjdarc_adapt(int bytes, arc_state_t *state) 3665168404Spjd{ 3666168404Spjd int mult; 3667208373Smm uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3668168404Spjd 3669185029Spjd if (state == arc_l2c_only) 3670185029Spjd return; 3671185029Spjd 3672168404Spjd ASSERT(bytes > 0); 3673168404Spjd /* 3674168404Spjd * Adapt the target size of the MRU list: 3675168404Spjd * - if we just hit in the MRU ghost list, then increase 3676168404Spjd * the target size of the MRU list. 3677168404Spjd * - if we just hit in the MFU ghost list, then increase 3678168404Spjd * the target size of the MFU list by decreasing the 3679168404Spjd * target size of the MRU list. 3680168404Spjd */ 3681168404Spjd if (state == arc_mru_ghost) { 3682168404Spjd mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 3683168404Spjd 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 3684209275Smm mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3685168404Spjd 3686208373Smm arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3687168404Spjd } else if (state == arc_mfu_ghost) { 3688208373Smm uint64_t delta; 3689208373Smm 3690168404Spjd mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 3691168404Spjd 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 3692209275Smm mult = MIN(mult, 10); 3693168404Spjd 3694208373Smm delta = MIN(bytes * mult, arc_p); 3695208373Smm arc_p = MAX(arc_p_min, arc_p - delta); 3696168404Spjd } 3697168404Spjd ASSERT((int64_t)arc_p >= 0); 3698168404Spjd 3699168404Spjd if (arc_reclaim_needed()) { 3700286763Smav cv_signal(&arc_reclaim_thread_cv); 3701168404Spjd return; 3702168404Spjd } 3703168404Spjd 3704168404Spjd if (arc_no_grow) 3705168404Spjd return; 3706168404Spjd 3707168404Spjd if (arc_c >= arc_c_max) 3708168404Spjd return; 3709168404Spjd 3710168404Spjd /* 3711168404Spjd * If we're within (2 * maxblocksize) bytes of the target 3712168404Spjd * cache size, increment the target cache size 3713168404Spjd */ 3714168404Spjd if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3715272483Ssmh DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3716168404Spjd atomic_add_64(&arc_c, (int64_t)bytes); 3717168404Spjd if (arc_c > arc_c_max) 3718168404Spjd arc_c = arc_c_max; 3719168404Spjd else if (state == arc_anon) 3720168404Spjd atomic_add_64(&arc_p, (int64_t)bytes); 3721168404Spjd if (arc_p > arc_c) 3722168404Spjd arc_p = arc_c; 3723168404Spjd } 3724168404Spjd ASSERT((int64_t)arc_p >= 0); 3725168404Spjd} 3726168404Spjd 3727168404Spjd/* 3728286763Smav * Check if arc_size has grown past our upper threshold, determined by 3729286763Smav * zfs_arc_overflow_shift. 3730168404Spjd */ 3731286763Smavstatic boolean_t 3732286763Smavarc_is_overflowing(void) 3733168404Spjd{ 3734286763Smav /* Always allow at least one block of overflow */ 3735286763Smav uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3736286763Smav arc_c >> zfs_arc_overflow_shift); 3737185029Spjd 3738286763Smav return (arc_size >= arc_c + overflow); 3739168404Spjd} 3740168404Spjd 3741168404Spjd/* 3742286763Smav * The buffer, supplied as the first argument, needs a data block. If we 3743286763Smav * are hitting the hard limit for the cache size, we must sleep, waiting 3744286763Smav * for the eviction thread to catch up. If we're past the target size 3745286763Smav * but below the hard limit, we'll only signal the reclaim thread and 3746286763Smav * continue on. 3747168404Spjd */ 3748168404Spjdstatic void 3749168404Spjdarc_get_data_buf(arc_buf_t *buf) 3750168404Spjd{ 3751286570Smav arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3752168404Spjd uint64_t size = buf->b_hdr->b_size; 3753286570Smav arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3754168404Spjd 3755168404Spjd arc_adapt(size, state); 3756168404Spjd 3757168404Spjd /* 3758286763Smav * If arc_size is currently overflowing, and has grown past our 3759286763Smav * upper limit, we must be adding data faster than the evict 3760286763Smav * thread can evict. Thus, to ensure we don't compound the 3761286763Smav * problem by adding more data and forcing arc_size to grow even 3762286763Smav * further past it's target size, we halt and wait for the 3763286763Smav * eviction thread to catch up. 3764286763Smav * 3765286763Smav * It's also possible that the reclaim thread is unable to evict 3766286763Smav * enough buffers to get arc_size below the overflow limit (e.g. 3767286763Smav * due to buffers being un-evictable, or hash lock collisions). 3768286763Smav * In this case, we want to proceed regardless if we're 3769286763Smav * overflowing; thus we don't use a while loop here. 3770168404Spjd */ 3771286763Smav if (arc_is_overflowing()) { 3772286763Smav mutex_enter(&arc_reclaim_lock); 3773286763Smav 3774286763Smav /* 3775286763Smav * Now that we've acquired the lock, we may no longer be 3776286763Smav * over the overflow limit, lets check. 3777286763Smav * 3778286763Smav * We're ignoring the case of spurious wake ups. If that 3779286763Smav * were to happen, it'd let this thread consume an ARC 3780286763Smav * buffer before it should have (i.e. before we're under 3781286763Smav * the overflow limit and were signalled by the reclaim 3782286763Smav * thread). As long as that is a rare occurrence, it 3783286763Smav * shouldn't cause any harm. 3784286763Smav */ 3785286763Smav if (arc_is_overflowing()) { 3786286763Smav cv_signal(&arc_reclaim_thread_cv); 3787286763Smav cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3788168404Spjd } 3789286763Smav 3790286763Smav mutex_exit(&arc_reclaim_lock); 3791168404Spjd } 3792168404Spjd 3793286763Smav if (type == ARC_BUFC_METADATA) { 3794286763Smav buf->b_data = zio_buf_alloc(size); 3795286763Smav arc_space_consume(size, ARC_SPACE_META); 3796168404Spjd } else { 3797286763Smav ASSERT(type == ARC_BUFC_DATA); 3798286763Smav buf->b_data = zio_data_buf_alloc(size); 3799286763Smav arc_space_consume(size, ARC_SPACE_DATA); 3800168404Spjd } 3801286763Smav 3802168404Spjd /* 3803168404Spjd * Update the state size. Note that ghost states have a 3804168404Spjd * "ghost size" and so don't need to be updated. 3805168404Spjd */ 3806286570Smav if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3807168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 3808168404Spjd 3809286570Smav atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size); 3810286763Smav 3811286763Smav /* 3812286763Smav * If this is reached via arc_read, the link is 3813286763Smav * protected by the hash lock. If reached via 3814286763Smav * arc_buf_alloc, the header should not be accessed by 3815286763Smav * any other thread. And, if reached via arc_read_done, 3816286763Smav * the hash lock will protect it if it's found in the 3817286763Smav * hash table; otherwise no other thread should be 3818286763Smav * trying to [add|remove]_reference it. 3819286763Smav */ 3820286763Smav if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3821286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3822286570Smav atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3823286570Smav size); 3824168404Spjd } 3825168404Spjd /* 3826168404Spjd * If we are growing the cache, and we are adding anonymous 3827168404Spjd * data, and we have outgrown arc_p, update arc_p 3828168404Spjd */ 3829286570Smav if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3830168404Spjd arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3831168404Spjd arc_p = MIN(arc_c, arc_p + size); 3832168404Spjd } 3833205231Skmacy ARCSTAT_BUMP(arcstat_allocated); 3834168404Spjd} 3835168404Spjd 3836168404Spjd/* 3837168404Spjd * This routine is called whenever a buffer is accessed. 3838168404Spjd * NOTE: the hash lock is dropped in this function. 3839168404Spjd */ 3840168404Spjdstatic void 3841275811Sdelphijarc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3842168404Spjd{ 3843219089Spjd clock_t now; 3844219089Spjd 3845168404Spjd ASSERT(MUTEX_HELD(hash_lock)); 3846286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 3847168404Spjd 3848286570Smav if (hdr->b_l1hdr.b_state == arc_anon) { 3849168404Spjd /* 3850168404Spjd * This buffer is not in the cache, and does not 3851168404Spjd * appear in our "ghost" list. Add the new buffer 3852168404Spjd * to the MRU state. 3853168404Spjd */ 3854168404Spjd 3855286570Smav ASSERT0(hdr->b_l1hdr.b_arc_access); 3856286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3857275811Sdelphij DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3858275811Sdelphij arc_change_state(arc_mru, hdr, hash_lock); 3859168404Spjd 3860286570Smav } else if (hdr->b_l1hdr.b_state == arc_mru) { 3861219089Spjd now = ddi_get_lbolt(); 3862219089Spjd 3863168404Spjd /* 3864168404Spjd * If this buffer is here because of a prefetch, then either: 3865168404Spjd * - clear the flag if this is a "referencing" read 3866168404Spjd * (any subsequent access will bump this into the MFU state). 3867168404Spjd * or 3868168404Spjd * - move the buffer to the head of the list if this is 3869168404Spjd * another prefetch (to make it less likely to be evicted). 3870168404Spjd */ 3871286570Smav if (HDR_PREFETCH(hdr)) { 3872286570Smav if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3873286763Smav /* link protected by hash lock */ 3874286763Smav ASSERT(multilist_link_active( 3875286570Smav &hdr->b_l1hdr.b_arc_node)); 3876168404Spjd } else { 3877275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3878168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 3879168404Spjd } 3880286570Smav hdr->b_l1hdr.b_arc_access = now; 3881168404Spjd return; 3882168404Spjd } 3883168404Spjd 3884168404Spjd /* 3885168404Spjd * This buffer has been "accessed" only once so far, 3886168404Spjd * but it is still in the cache. Move it to the MFU 3887168404Spjd * state. 3888168404Spjd */ 3889286570Smav if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3890168404Spjd /* 3891168404Spjd * More than 125ms have passed since we 3892168404Spjd * instantiated this buffer. Move it to the 3893168404Spjd * most frequently used state. 3894168404Spjd */ 3895286570Smav hdr->b_l1hdr.b_arc_access = now; 3896275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3897275811Sdelphij arc_change_state(arc_mfu, hdr, hash_lock); 3898168404Spjd } 3899168404Spjd ARCSTAT_BUMP(arcstat_mru_hits); 3900286570Smav } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3901168404Spjd arc_state_t *new_state; 3902168404Spjd /* 3903168404Spjd * This buffer has been "accessed" recently, but 3904168404Spjd * was evicted from the cache. Move it to the 3905168404Spjd * MFU state. 3906168404Spjd */ 3907168404Spjd 3908286570Smav if (HDR_PREFETCH(hdr)) { 3909168404Spjd new_state = arc_mru; 3910286570Smav if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3911275811Sdelphij hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3912275811Sdelphij DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3913168404Spjd } else { 3914168404Spjd new_state = arc_mfu; 3915275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3916168404Spjd } 3917168404Spjd 3918286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3919275811Sdelphij arc_change_state(new_state, hdr, hash_lock); 3920168404Spjd 3921168404Spjd ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3922286570Smav } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3923168404Spjd /* 3924168404Spjd * This buffer has been accessed more than once and is 3925168404Spjd * still in the cache. Keep it in the MFU state. 3926168404Spjd * 3927168404Spjd * NOTE: an add_reference() that occurred when we did 3928168404Spjd * the arc_read() will have kicked this off the list. 3929168404Spjd * If it was a prefetch, we will explicitly move it to 3930168404Spjd * the head of the list now. 3931168404Spjd */ 3932286570Smav if ((HDR_PREFETCH(hdr)) != 0) { 3933286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3934286763Smav /* link protected by hash_lock */ 3935286763Smav ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3936168404Spjd } 3937168404Spjd ARCSTAT_BUMP(arcstat_mfu_hits); 3938286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3939286570Smav } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 3940168404Spjd arc_state_t *new_state = arc_mfu; 3941168404Spjd /* 3942168404Spjd * This buffer has been accessed more than once but has 3943168404Spjd * been evicted from the cache. Move it back to the 3944168404Spjd * MFU state. 3945168404Spjd */ 3946168404Spjd 3947286570Smav if (HDR_PREFETCH(hdr)) { 3948168404Spjd /* 3949168404Spjd * This is a prefetch access... 3950168404Spjd * move this block back to the MRU state. 3951168404Spjd */ 3952286570Smav ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3953168404Spjd new_state = arc_mru; 3954168404Spjd } 3955168404Spjd 3956286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3957275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3958275811Sdelphij arc_change_state(new_state, hdr, hash_lock); 3959168404Spjd 3960168404Spjd ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3961286570Smav } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 3962185029Spjd /* 3963185029Spjd * This buffer is on the 2nd Level ARC. 3964185029Spjd */ 3965185029Spjd 3966286570Smav hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3967275811Sdelphij DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3968275811Sdelphij arc_change_state(arc_mfu, hdr, hash_lock); 3969168404Spjd } else { 3970168404Spjd ASSERT(!"invalid arc state"); 3971168404Spjd } 3972168404Spjd} 3973168404Spjd 3974168404Spjd/* a generic arc_done_func_t which you can use */ 3975168404Spjd/* ARGSUSED */ 3976168404Spjdvoid 3977168404Spjdarc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3978168404Spjd{ 3979219089Spjd if (zio == NULL || zio->io_error == 0) 3980219089Spjd bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3981248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 3982168404Spjd} 3983168404Spjd 3984185029Spjd/* a generic arc_done_func_t */ 3985168404Spjdvoid 3986168404Spjdarc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3987168404Spjd{ 3988168404Spjd arc_buf_t **bufp = arg; 3989168404Spjd if (zio && zio->io_error) { 3990248571Smm VERIFY(arc_buf_remove_ref(buf, arg)); 3991168404Spjd *bufp = NULL; 3992168404Spjd } else { 3993168404Spjd *bufp = buf; 3994219089Spjd ASSERT(buf->b_data); 3995168404Spjd } 3996168404Spjd} 3997168404Spjd 3998168404Spjdstatic void 3999168404Spjdarc_read_done(zio_t *zio) 4000168404Spjd{ 4001268075Sdelphij arc_buf_hdr_t *hdr; 4002168404Spjd arc_buf_t *buf; 4003168404Spjd arc_buf_t *abuf; /* buffer we're assigning to callback */ 4004268075Sdelphij kmutex_t *hash_lock = NULL; 4005168404Spjd arc_callback_t *callback_list, *acb; 4006168404Spjd int freeable = FALSE; 4007168404Spjd 4008168404Spjd buf = zio->io_private; 4009168404Spjd hdr = buf->b_hdr; 4010168404Spjd 4011168404Spjd /* 4012168404Spjd * The hdr was inserted into hash-table and removed from lists 4013168404Spjd * prior to starting I/O. We should find this header, since 4014168404Spjd * it's in the hash table, and it should be legit since it's 4015168404Spjd * not possible to evict it during the I/O. The only possible 4016168404Spjd * reason for it not to be found is if we were freed during the 4017168404Spjd * read. 4018168404Spjd */ 4019268075Sdelphij if (HDR_IN_HASH_TABLE(hdr)) { 4020268075Sdelphij ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4021268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[0], ==, 4022268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[0]); 4023268075Sdelphij ASSERT3U(hdr->b_dva.dva_word[1], ==, 4024268075Sdelphij BP_IDENTITY(zio->io_bp)->dva_word[1]); 4025168404Spjd 4026268075Sdelphij arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4027268075Sdelphij &hash_lock); 4028168404Spjd 4029268075Sdelphij ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4030268075Sdelphij hash_lock == NULL) || 4031268075Sdelphij (found == hdr && 4032268075Sdelphij DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4033268075Sdelphij (found == hdr && HDR_L2_READING(hdr))); 4034268075Sdelphij } 4035268075Sdelphij 4036275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4037286570Smav if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4038275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4039206796Spjd 4040168404Spjd /* byteswap if necessary */ 4041286570Smav callback_list = hdr->b_l1hdr.b_acb; 4042168404Spjd ASSERT(callback_list != NULL); 4043209101Smm if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4044236884Smm dmu_object_byteswap_t bswap = 4045236884Smm DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4046185029Spjd arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4047185029Spjd byteswap_uint64_array : 4048236884Smm dmu_ot_byteswap[bswap].ob_func; 4049185029Spjd func(buf->b_data, hdr->b_size); 4050185029Spjd } 4051168404Spjd 4052185029Spjd arc_cksum_compute(buf, B_FALSE); 4053240133Smm#ifdef illumos 4054240133Smm arc_buf_watch(buf); 4055277300Ssmh#endif 4056168404Spjd 4057286570Smav if (hash_lock && zio->io_error == 0 && 4058286570Smav hdr->b_l1hdr.b_state == arc_anon) { 4059219089Spjd /* 4060219089Spjd * Only call arc_access on anonymous buffers. This is because 4061219089Spjd * if we've issued an I/O for an evicted buffer, we've already 4062219089Spjd * called arc_access (to prevent any simultaneous readers from 4063219089Spjd * getting confused). 4064219089Spjd */ 4065219089Spjd arc_access(hdr, hash_lock); 4066219089Spjd } 4067219089Spjd 4068168404Spjd /* create copies of the data buffer for the callers */ 4069168404Spjd abuf = buf; 4070168404Spjd for (acb = callback_list; acb; acb = acb->acb_next) { 4071168404Spjd if (acb->acb_done) { 4072242845Sdelphij if (abuf == NULL) { 4073242845Sdelphij ARCSTAT_BUMP(arcstat_duplicate_reads); 4074168404Spjd abuf = arc_buf_clone(buf); 4075242845Sdelphij } 4076168404Spjd acb->acb_buf = abuf; 4077168404Spjd abuf = NULL; 4078168404Spjd } 4079168404Spjd } 4080286570Smav hdr->b_l1hdr.b_acb = NULL; 4081275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4082168404Spjd ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4083219089Spjd if (abuf == buf) { 4084219089Spjd ASSERT(buf->b_efunc == NULL); 4085286570Smav ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4086275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4087219089Spjd } 4088168404Spjd 4089286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4090286570Smav callback_list != NULL); 4091168404Spjd 4092168404Spjd if (zio->io_error != 0) { 4093275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_ERROR; 4094286570Smav if (hdr->b_l1hdr.b_state != arc_anon) 4095168404Spjd arc_change_state(arc_anon, hdr, hash_lock); 4096168404Spjd if (HDR_IN_HASH_TABLE(hdr)) 4097168404Spjd buf_hash_remove(hdr); 4098286570Smav freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4099168404Spjd } 4100168404Spjd 4101168404Spjd /* 4102168404Spjd * Broadcast before we drop the hash_lock to avoid the possibility 4103168404Spjd * that the hdr (and hence the cv) might be freed before we get to 4104168404Spjd * the cv_broadcast(). 4105168404Spjd */ 4106286570Smav cv_broadcast(&hdr->b_l1hdr.b_cv); 4107168404Spjd 4108286570Smav if (hash_lock != NULL) { 4109168404Spjd mutex_exit(hash_lock); 4110168404Spjd } else { 4111168404Spjd /* 4112168404Spjd * This block was freed while we waited for the read to 4113168404Spjd * complete. It has been removed from the hash table and 4114168404Spjd * moved to the anonymous state (so that it won't show up 4115168404Spjd * in the cache). 4116168404Spjd */ 4117286570Smav ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4118286570Smav freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4119168404Spjd } 4120168404Spjd 4121168404Spjd /* execute each callback and free its structure */ 4122168404Spjd while ((acb = callback_list) != NULL) { 4123168404Spjd if (acb->acb_done) 4124168404Spjd acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4125168404Spjd 4126168404Spjd if (acb->acb_zio_dummy != NULL) { 4127168404Spjd acb->acb_zio_dummy->io_error = zio->io_error; 4128168404Spjd zio_nowait(acb->acb_zio_dummy); 4129168404Spjd } 4130168404Spjd 4131168404Spjd callback_list = acb->acb_next; 4132168404Spjd kmem_free(acb, sizeof (arc_callback_t)); 4133168404Spjd } 4134168404Spjd 4135168404Spjd if (freeable) 4136168404Spjd arc_hdr_destroy(hdr); 4137168404Spjd} 4138168404Spjd 4139168404Spjd/* 4140286762Smav * "Read" the block at the specified DVA (in bp) via the 4141168404Spjd * cache. If the block is found in the cache, invoke the provided 4142168404Spjd * callback immediately and return. Note that the `zio' parameter 4143168404Spjd * in the callback will be NULL in this case, since no IO was 4144168404Spjd * required. If the block is not in the cache pass the read request 4145168404Spjd * on to the spa with a substitute callback function, so that the 4146168404Spjd * requested block will be added to the cache. 4147168404Spjd * 4148168404Spjd * If a read request arrives for a block that has a read in-progress, 4149168404Spjd * either wait for the in-progress read to complete (and return the 4150168404Spjd * results); or, if this is a read with a "done" func, add a record 4151168404Spjd * to the read to invoke the "done" func when the read completes, 4152168404Spjd * and return; or just return. 4153168404Spjd * 4154168404Spjd * arc_read_done() will invoke all the requested "done" functions 4155168404Spjd * for readers of this block. 4156168404Spjd */ 4157168404Spjdint 4158246666Smmarc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4159275811Sdelphij void *private, zio_priority_t priority, int zio_flags, 4160275811Sdelphij arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4161168404Spjd{ 4162268075Sdelphij arc_buf_hdr_t *hdr = NULL; 4163247187Smm arc_buf_t *buf = NULL; 4164268075Sdelphij kmutex_t *hash_lock = NULL; 4165185029Spjd zio_t *rzio; 4166228103Smm uint64_t guid = spa_load_guid(spa); 4167168404Spjd 4168268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp) || 4169268075Sdelphij BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4170268075Sdelphij 4171168404Spjdtop: 4172268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 4173268075Sdelphij /* 4174268075Sdelphij * Embedded BP's have no DVA and require no I/O to "read". 4175268075Sdelphij * Create an anonymous arc buf to back it. 4176268075Sdelphij */ 4177268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 4178268075Sdelphij } 4179168404Spjd 4180286570Smav if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4181268075Sdelphij 4182275811Sdelphij *arc_flags |= ARC_FLAG_CACHED; 4183168404Spjd 4184168404Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 4185168404Spjd 4186275811Sdelphij if (*arc_flags & ARC_FLAG_WAIT) { 4187286570Smav cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4188168404Spjd mutex_exit(hash_lock); 4189168404Spjd goto top; 4190168404Spjd } 4191275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4192168404Spjd 4193168404Spjd if (done) { 4194168404Spjd arc_callback_t *acb = NULL; 4195168404Spjd 4196168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), 4197168404Spjd KM_SLEEP); 4198168404Spjd acb->acb_done = done; 4199168404Spjd acb->acb_private = private; 4200168404Spjd if (pio != NULL) 4201168404Spjd acb->acb_zio_dummy = zio_null(pio, 4202209962Smm spa, NULL, NULL, NULL, zio_flags); 4203168404Spjd 4204168404Spjd ASSERT(acb->acb_done != NULL); 4205286570Smav acb->acb_next = hdr->b_l1hdr.b_acb; 4206286570Smav hdr->b_l1hdr.b_acb = acb; 4207168404Spjd add_reference(hdr, hash_lock, private); 4208168404Spjd mutex_exit(hash_lock); 4209168404Spjd return (0); 4210168404Spjd } 4211168404Spjd mutex_exit(hash_lock); 4212168404Spjd return (0); 4213168404Spjd } 4214168404Spjd 4215286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4216286570Smav hdr->b_l1hdr.b_state == arc_mfu); 4217168404Spjd 4218168404Spjd if (done) { 4219168404Spjd add_reference(hdr, hash_lock, private); 4220168404Spjd /* 4221168404Spjd * If this block is already in use, create a new 4222168404Spjd * copy of the data so that we will be guaranteed 4223168404Spjd * that arc_release() will always succeed. 4224168404Spjd */ 4225286570Smav buf = hdr->b_l1hdr.b_buf; 4226168404Spjd ASSERT(buf); 4227168404Spjd ASSERT(buf->b_data); 4228168404Spjd if (HDR_BUF_AVAILABLE(hdr)) { 4229168404Spjd ASSERT(buf->b_efunc == NULL); 4230275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4231168404Spjd } else { 4232168404Spjd buf = arc_buf_clone(buf); 4233168404Spjd } 4234219089Spjd 4235275811Sdelphij } else if (*arc_flags & ARC_FLAG_PREFETCH && 4236286570Smav refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4237275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 4238168404Spjd } 4239168404Spjd DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4240168404Spjd arc_access(hdr, hash_lock); 4241275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 4242275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 4243275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 4244275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4245168404Spjd mutex_exit(hash_lock); 4246168404Spjd ARCSTAT_BUMP(arcstat_hits); 4247286570Smav ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4248286570Smav demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4249168404Spjd data, metadata, hits); 4250168404Spjd 4251168404Spjd if (done) 4252168404Spjd done(NULL, buf, private); 4253168404Spjd } else { 4254168404Spjd uint64_t size = BP_GET_LSIZE(bp); 4255268075Sdelphij arc_callback_t *acb; 4256185029Spjd vdev_t *vd = NULL; 4257247187Smm uint64_t addr = 0; 4258208373Smm boolean_t devw = B_FALSE; 4259258389Savg enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4260286570Smav int32_t b_asize = 0; 4261168404Spjd 4262168404Spjd if (hdr == NULL) { 4263168404Spjd /* this block is not in the cache */ 4264268075Sdelphij arc_buf_hdr_t *exists = NULL; 4265168404Spjd arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4266168404Spjd buf = arc_buf_alloc(spa, size, private, type); 4267168404Spjd hdr = buf->b_hdr; 4268268075Sdelphij if (!BP_IS_EMBEDDED(bp)) { 4269268075Sdelphij hdr->b_dva = *BP_IDENTITY(bp); 4270268075Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4271268075Sdelphij exists = buf_hash_insert(hdr, &hash_lock); 4272268075Sdelphij } 4273268075Sdelphij if (exists != NULL) { 4274168404Spjd /* somebody beat us to the hash insert */ 4275168404Spjd mutex_exit(hash_lock); 4276219089Spjd buf_discard_identity(hdr); 4277168404Spjd (void) arc_buf_remove_ref(buf, private); 4278168404Spjd goto top; /* restart the IO request */ 4279168404Spjd } 4280275811Sdelphij 4281168404Spjd /* if this is a prefetch, we don't have a reference */ 4282275811Sdelphij if (*arc_flags & ARC_FLAG_PREFETCH) { 4283168404Spjd (void) remove_reference(hdr, hash_lock, 4284168404Spjd private); 4285275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 4286168404Spjd } 4287275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 4288275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 4289275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 4290275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4291168404Spjd if (BP_GET_LEVEL(bp) > 0) 4292275811Sdelphij hdr->b_flags |= ARC_FLAG_INDIRECT; 4293168404Spjd } else { 4294286570Smav /* 4295286570Smav * This block is in the ghost cache. If it was L2-only 4296286570Smav * (and thus didn't have an L1 hdr), we realloc the 4297286570Smav * header to add an L1 hdr. 4298286570Smav */ 4299286570Smav if (!HDR_HAS_L1HDR(hdr)) { 4300286570Smav hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4301286570Smav hdr_full_cache); 4302286570Smav } 4303286570Smav 4304286570Smav ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4305168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4306286570Smav ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4307286763Smav ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4308168404Spjd 4309168404Spjd /* if this is a prefetch, we don't have a reference */ 4310275811Sdelphij if (*arc_flags & ARC_FLAG_PREFETCH) 4311275811Sdelphij hdr->b_flags |= ARC_FLAG_PREFETCH; 4312168404Spjd else 4313168404Spjd add_reference(hdr, hash_lock, private); 4314275811Sdelphij if (*arc_flags & ARC_FLAG_L2CACHE) 4315275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 4316275811Sdelphij if (*arc_flags & ARC_FLAG_L2COMPRESS) 4317275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4318185029Spjd buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4319168404Spjd buf->b_hdr = hdr; 4320168404Spjd buf->b_data = NULL; 4321168404Spjd buf->b_efunc = NULL; 4322168404Spjd buf->b_private = NULL; 4323168404Spjd buf->b_next = NULL; 4324286570Smav hdr->b_l1hdr.b_buf = buf; 4325286570Smav ASSERT0(hdr->b_l1hdr.b_datacnt); 4326286570Smav hdr->b_l1hdr.b_datacnt = 1; 4327219089Spjd arc_get_data_buf(buf); 4328219089Spjd arc_access(hdr, hash_lock); 4329168404Spjd } 4330168404Spjd 4331286570Smav ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4332219089Spjd 4333168404Spjd acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4334168404Spjd acb->acb_done = done; 4335168404Spjd acb->acb_private = private; 4336168404Spjd 4337286570Smav ASSERT(hdr->b_l1hdr.b_acb == NULL); 4338286570Smav hdr->b_l1hdr.b_acb = acb; 4339275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4340168404Spjd 4341286570Smav if (HDR_HAS_L2HDR(hdr) && 4342286570Smav (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4343286570Smav devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4344286570Smav addr = hdr->b_l2hdr.b_daddr; 4345286570Smav b_compress = HDR_GET_COMPRESS(hdr); 4346286570Smav b_asize = hdr->b_l2hdr.b_asize; 4347185029Spjd /* 4348185029Spjd * Lock out device removal. 4349185029Spjd */ 4350185029Spjd if (vdev_is_dead(vd) || 4351185029Spjd !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4352185029Spjd vd = NULL; 4353185029Spjd } 4354185029Spjd 4355268075Sdelphij if (hash_lock != NULL) 4356268075Sdelphij mutex_exit(hash_lock); 4357168404Spjd 4358251629Sdelphij /* 4359251629Sdelphij * At this point, we have a level 1 cache miss. Try again in 4360251629Sdelphij * L2ARC if possible. 4361251629Sdelphij */ 4362168404Spjd ASSERT3U(hdr->b_size, ==, size); 4363219089Spjd DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4364268123Sdelphij uint64_t, size, zbookmark_phys_t *, zb); 4365168404Spjd ARCSTAT_BUMP(arcstat_misses); 4366286570Smav ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4367286570Smav demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4368168404Spjd data, metadata, misses); 4369228392Spjd#ifdef _KERNEL 4370228392Spjd curthread->td_ru.ru_inblock++; 4371228392Spjd#endif 4372168404Spjd 4373208373Smm if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4374185029Spjd /* 4375185029Spjd * Read from the L2ARC if the following are true: 4376185029Spjd * 1. The L2ARC vdev was previously cached. 4377185029Spjd * 2. This buffer still has L2ARC metadata. 4378185029Spjd * 3. This buffer isn't currently writing to the L2ARC. 4379185029Spjd * 4. The L2ARC entry wasn't evicted, which may 4380185029Spjd * also have invalidated the vdev. 4381208373Smm * 5. This isn't prefetch and l2arc_noprefetch is set. 4382185029Spjd */ 4383286570Smav if (HDR_HAS_L2HDR(hdr) && 4384208373Smm !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4385208373Smm !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4386185029Spjd l2arc_read_callback_t *cb; 4387185029Spjd 4388185029Spjd DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4389185029Spjd ARCSTAT_BUMP(arcstat_l2_hits); 4390185029Spjd 4391185029Spjd cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4392185029Spjd KM_SLEEP); 4393185029Spjd cb->l2rcb_buf = buf; 4394185029Spjd cb->l2rcb_spa = spa; 4395185029Spjd cb->l2rcb_bp = *bp; 4396185029Spjd cb->l2rcb_zb = *zb; 4397185029Spjd cb->l2rcb_flags = zio_flags; 4398258389Savg cb->l2rcb_compress = b_compress; 4399185029Spjd 4400247187Smm ASSERT(addr >= VDEV_LABEL_START_SIZE && 4401247187Smm addr + size < vd->vdev_psize - 4402247187Smm VDEV_LABEL_END_SIZE); 4403247187Smm 4404185029Spjd /* 4405185029Spjd * l2arc read. The SCL_L2ARC lock will be 4406185029Spjd * released by l2arc_read_done(). 4407251478Sdelphij * Issue a null zio if the underlying buffer 4408251478Sdelphij * was squashed to zero size by compression. 4409185029Spjd */ 4410258389Savg if (b_compress == ZIO_COMPRESS_EMPTY) { 4411251478Sdelphij rzio = zio_null(pio, spa, vd, 4412251478Sdelphij l2arc_read_done, cb, 4413251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 4414251478Sdelphij ZIO_FLAG_CANFAIL | 4415251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 4416251478Sdelphij ZIO_FLAG_DONT_RETRY); 4417251478Sdelphij } else { 4418251478Sdelphij rzio = zio_read_phys(pio, vd, addr, 4419258389Savg b_asize, buf->b_data, 4420258389Savg ZIO_CHECKSUM_OFF, 4421251478Sdelphij l2arc_read_done, cb, priority, 4422251478Sdelphij zio_flags | ZIO_FLAG_DONT_CACHE | 4423251478Sdelphij ZIO_FLAG_CANFAIL | 4424251478Sdelphij ZIO_FLAG_DONT_PROPAGATE | 4425251478Sdelphij ZIO_FLAG_DONT_RETRY, B_FALSE); 4426251478Sdelphij } 4427185029Spjd DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4428185029Spjd zio_t *, rzio); 4429258389Savg ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4430185029Spjd 4431275811Sdelphij if (*arc_flags & ARC_FLAG_NOWAIT) { 4432185029Spjd zio_nowait(rzio); 4433185029Spjd return (0); 4434185029Spjd } 4435185029Spjd 4436275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_WAIT); 4437185029Spjd if (zio_wait(rzio) == 0) 4438185029Spjd return (0); 4439185029Spjd 4440185029Spjd /* l2arc read error; goto zio_read() */ 4441185029Spjd } else { 4442185029Spjd DTRACE_PROBE1(l2arc__miss, 4443185029Spjd arc_buf_hdr_t *, hdr); 4444185029Spjd ARCSTAT_BUMP(arcstat_l2_misses); 4445185029Spjd if (HDR_L2_WRITING(hdr)) 4446185029Spjd ARCSTAT_BUMP(arcstat_l2_rw_clash); 4447185029Spjd spa_config_exit(spa, SCL_L2ARC, vd); 4448185029Spjd } 4449208373Smm } else { 4450208373Smm if (vd != NULL) 4451208373Smm spa_config_exit(spa, SCL_L2ARC, vd); 4452208373Smm if (l2arc_ndev != 0) { 4453208373Smm DTRACE_PROBE1(l2arc__miss, 4454208373Smm arc_buf_hdr_t *, hdr); 4455208373Smm ARCSTAT_BUMP(arcstat_l2_misses); 4456208373Smm } 4457185029Spjd } 4458185029Spjd 4459168404Spjd rzio = zio_read(pio, spa, bp, buf->b_data, size, 4460185029Spjd arc_read_done, buf, priority, zio_flags, zb); 4461168404Spjd 4462275811Sdelphij if (*arc_flags & ARC_FLAG_WAIT) 4463168404Spjd return (zio_wait(rzio)); 4464168404Spjd 4465275811Sdelphij ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4466168404Spjd zio_nowait(rzio); 4467168404Spjd } 4468168404Spjd return (0); 4469168404Spjd} 4470168404Spjd 4471168404Spjdvoid 4472168404Spjdarc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4473168404Spjd{ 4474168404Spjd ASSERT(buf->b_hdr != NULL); 4475286570Smav ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4476286570Smav ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4477286570Smav func == NULL); 4478219089Spjd ASSERT(buf->b_efunc == NULL); 4479219089Spjd ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4480219089Spjd 4481168404Spjd buf->b_efunc = func; 4482168404Spjd buf->b_private = private; 4483168404Spjd} 4484168404Spjd 4485168404Spjd/* 4486251520Sdelphij * Notify the arc that a block was freed, and thus will never be used again. 4487251520Sdelphij */ 4488251520Sdelphijvoid 4489251520Sdelphijarc_freed(spa_t *spa, const blkptr_t *bp) 4490251520Sdelphij{ 4491251520Sdelphij arc_buf_hdr_t *hdr; 4492251520Sdelphij kmutex_t *hash_lock; 4493251520Sdelphij uint64_t guid = spa_load_guid(spa); 4494251520Sdelphij 4495268075Sdelphij ASSERT(!BP_IS_EMBEDDED(bp)); 4496268075Sdelphij 4497268075Sdelphij hdr = buf_hash_find(guid, bp, &hash_lock); 4498251520Sdelphij if (hdr == NULL) 4499251520Sdelphij return; 4500251520Sdelphij if (HDR_BUF_AVAILABLE(hdr)) { 4501286570Smav arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4502251520Sdelphij add_reference(hdr, hash_lock, FTAG); 4503275811Sdelphij hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4504251520Sdelphij mutex_exit(hash_lock); 4505251520Sdelphij 4506251520Sdelphij arc_release(buf, FTAG); 4507251520Sdelphij (void) arc_buf_remove_ref(buf, FTAG); 4508251520Sdelphij } else { 4509251520Sdelphij mutex_exit(hash_lock); 4510251520Sdelphij } 4511251520Sdelphij 4512251520Sdelphij} 4513251520Sdelphij 4514251520Sdelphij/* 4515268858Sdelphij * Clear the user eviction callback set by arc_set_callback(), first calling 4516268858Sdelphij * it if it exists. Because the presence of a callback keeps an arc_buf cached 4517268858Sdelphij * clearing the callback may result in the arc_buf being destroyed. However, 4518268858Sdelphij * it will not result in the *last* arc_buf being destroyed, hence the data 4519268858Sdelphij * will remain cached in the ARC. We make a copy of the arc buffer here so 4520268858Sdelphij * that we can process the callback without holding any locks. 4521268858Sdelphij * 4522268858Sdelphij * It's possible that the callback is already in the process of being cleared 4523268858Sdelphij * by another thread. In this case we can not clear the callback. 4524268858Sdelphij * 4525268858Sdelphij * Returns B_TRUE if the callback was successfully called and cleared. 4526168404Spjd */ 4527268858Sdelphijboolean_t 4528268858Sdelphijarc_clear_callback(arc_buf_t *buf) 4529168404Spjd{ 4530168404Spjd arc_buf_hdr_t *hdr; 4531168404Spjd kmutex_t *hash_lock; 4532268858Sdelphij arc_evict_func_t *efunc = buf->b_efunc; 4533268858Sdelphij void *private = buf->b_private; 4534206796Spjd 4535219089Spjd mutex_enter(&buf->b_evict_lock); 4536168404Spjd hdr = buf->b_hdr; 4537168404Spjd if (hdr == NULL) { 4538168404Spjd /* 4539168404Spjd * We are in arc_do_user_evicts(). 4540168404Spjd */ 4541168404Spjd ASSERT(buf->b_data == NULL); 4542219089Spjd mutex_exit(&buf->b_evict_lock); 4543268858Sdelphij return (B_FALSE); 4544185029Spjd } else if (buf->b_data == NULL) { 4545185029Spjd /* 4546185029Spjd * We are on the eviction list; process this buffer now 4547185029Spjd * but let arc_do_user_evicts() do the reaping. 4548185029Spjd */ 4549185029Spjd buf->b_efunc = NULL; 4550219089Spjd mutex_exit(&buf->b_evict_lock); 4551268858Sdelphij VERIFY0(efunc(private)); 4552268858Sdelphij return (B_TRUE); 4553168404Spjd } 4554168404Spjd hash_lock = HDR_LOCK(hdr); 4555168404Spjd mutex_enter(hash_lock); 4556219089Spjd hdr = buf->b_hdr; 4557219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4558168404Spjd 4559286570Smav ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4560286570Smav hdr->b_l1hdr.b_datacnt); 4561286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4562286570Smav hdr->b_l1hdr.b_state == arc_mfu); 4563168404Spjd 4564268858Sdelphij buf->b_efunc = NULL; 4565268858Sdelphij buf->b_private = NULL; 4566168404Spjd 4567286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 4568268858Sdelphij mutex_exit(&buf->b_evict_lock); 4569286763Smav arc_buf_destroy(buf, TRUE); 4570268858Sdelphij } else { 4571286570Smav ASSERT(buf == hdr->b_l1hdr.b_buf); 4572275811Sdelphij hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4573268858Sdelphij mutex_exit(&buf->b_evict_lock); 4574268858Sdelphij } 4575168404Spjd 4576168404Spjd mutex_exit(hash_lock); 4577268858Sdelphij VERIFY0(efunc(private)); 4578268858Sdelphij return (B_TRUE); 4579168404Spjd} 4580168404Spjd 4581168404Spjd/* 4582251629Sdelphij * Release this buffer from the cache, making it an anonymous buffer. This 4583251629Sdelphij * must be done after a read and prior to modifying the buffer contents. 4584168404Spjd * If the buffer has more than one reference, we must make 4585185029Spjd * a new hdr for the buffer. 4586168404Spjd */ 4587168404Spjdvoid 4588168404Spjdarc_release(arc_buf_t *buf, void *tag) 4589168404Spjd{ 4590286570Smav arc_buf_hdr_t *hdr = buf->b_hdr; 4591168404Spjd 4592286763Smav ASSERT(HDR_HAS_L1HDR(hdr)); 4593286763Smav 4594219089Spjd /* 4595219089Spjd * It would be nice to assert that if it's DMU metadata (level > 4596219089Spjd * 0 || it's the dnode file), then it must be syncing context. 4597219089Spjd * But we don't know that information at this level. 4598219089Spjd */ 4599219089Spjd 4600219089Spjd mutex_enter(&buf->b_evict_lock); 4601286570Smav /* 4602286570Smav * We don't grab the hash lock prior to this check, because if 4603286570Smav * the buffer's header is in the arc_anon state, it won't be 4604286570Smav * linked into the hash table. 4605286570Smav */ 4606286570Smav if (hdr->b_l1hdr.b_state == arc_anon) { 4607286570Smav mutex_exit(&buf->b_evict_lock); 4608286570Smav ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4609286570Smav ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4610286570Smav ASSERT(!HDR_HAS_L2HDR(hdr)); 4611286570Smav ASSERT(BUF_EMPTY(hdr)); 4612286570Smav ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4613286570Smav ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4614286570Smav ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4615185029Spjd 4616286570Smav ASSERT3P(buf->b_efunc, ==, NULL); 4617286570Smav ASSERT3P(buf->b_private, ==, NULL); 4618168404Spjd 4619286570Smav hdr->b_l1hdr.b_arc_access = 0; 4620286570Smav arc_buf_thaw(buf); 4621286570Smav 4622286570Smav return; 4623168404Spjd } 4624168404Spjd 4625286570Smav kmutex_t *hash_lock = HDR_LOCK(hdr); 4626286570Smav mutex_enter(hash_lock); 4627286570Smav 4628286570Smav /* 4629286570Smav * This assignment is only valid as long as the hash_lock is 4630286570Smav * held, we must be careful not to reference state or the 4631286570Smav * b_state field after dropping the lock. 4632286570Smav */ 4633286570Smav arc_state_t *state = hdr->b_l1hdr.b_state; 4634286570Smav ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4635286570Smav ASSERT3P(state, !=, arc_anon); 4636286570Smav 4637286570Smav /* this buffer is not on any list */ 4638286570Smav ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4639286570Smav 4640286570Smav if (HDR_HAS_L2HDR(hdr)) { 4641286570Smav mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4642286570Smav 4643286570Smav /* 4644286598Smav * We have to recheck this conditional again now that 4645286598Smav * we're holding the l2ad_mtx to prevent a race with 4646286598Smav * another thread which might be concurrently calling 4647286598Smav * l2arc_evict(). In that case, l2arc_evict() might have 4648286598Smav * destroyed the header's L2 portion as we were waiting 4649286598Smav * to acquire the l2ad_mtx. 4650286570Smav */ 4651286598Smav if (HDR_HAS_L2HDR(hdr)) { 4652286647Smav if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4653286647Smav trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4654286647Smav hdr->b_l2hdr.b_daddr, 4655286647Smav hdr->b_l2hdr.b_asize, 0); 4656286598Smav arc_hdr_l2hdr_destroy(hdr); 4657286598Smav } 4658286570Smav 4659286570Smav mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4660185029Spjd } 4661185029Spjd 4662168404Spjd /* 4663168404Spjd * Do we have more than one buf? 4664168404Spjd */ 4665286570Smav if (hdr->b_l1hdr.b_datacnt > 1) { 4666168404Spjd arc_buf_hdr_t *nhdr; 4667168404Spjd arc_buf_t **bufp; 4668168404Spjd uint64_t blksz = hdr->b_size; 4669209962Smm uint64_t spa = hdr->b_spa; 4670286570Smav arc_buf_contents_t type = arc_buf_type(hdr); 4671185029Spjd uint32_t flags = hdr->b_flags; 4672168404Spjd 4673286570Smav ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4674168404Spjd /* 4675219089Spjd * Pull the data off of this hdr and attach it to 4676219089Spjd * a new anonymous hdr. 4677168404Spjd */ 4678168404Spjd (void) remove_reference(hdr, hash_lock, tag); 4679286570Smav bufp = &hdr->b_l1hdr.b_buf; 4680168404Spjd while (*bufp != buf) 4681168404Spjd bufp = &(*bufp)->b_next; 4682219089Spjd *bufp = buf->b_next; 4683168404Spjd buf->b_next = NULL; 4684168404Spjd 4685286570Smav ASSERT3P(state, !=, arc_l2c_only); 4686286570Smav ASSERT3U(state->arcs_size, >=, hdr->b_size); 4687286570Smav atomic_add_64(&state->arcs_size, -hdr->b_size); 4688286570Smav if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4689286570Smav ASSERT3P(state, !=, arc_l2c_only); 4690286570Smav uint64_t *size = &state->arcs_lsize[type]; 4691185029Spjd ASSERT3U(*size, >=, hdr->b_size); 4692185029Spjd atomic_add_64(size, -hdr->b_size); 4693168404Spjd } 4694242845Sdelphij 4695242845Sdelphij /* 4696242845Sdelphij * We're releasing a duplicate user data buffer, update 4697242845Sdelphij * our statistics accordingly. 4698242845Sdelphij */ 4699286570Smav if (HDR_ISTYPE_DATA(hdr)) { 4700242845Sdelphij ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4701242845Sdelphij ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4702242845Sdelphij -hdr->b_size); 4703242845Sdelphij } 4704286570Smav hdr->b_l1hdr.b_datacnt -= 1; 4705168404Spjd arc_cksum_verify(buf); 4706240133Smm#ifdef illumos 4707240133Smm arc_buf_unwatch(buf); 4708277300Ssmh#endif 4709168404Spjd 4710168404Spjd mutex_exit(hash_lock); 4711168404Spjd 4712286570Smav nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4713168404Spjd nhdr->b_size = blksz; 4714168404Spjd nhdr->b_spa = spa; 4715286570Smav 4716275811Sdelphij nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4717286570Smav nhdr->b_flags |= arc_bufc_to_flags(type); 4718286570Smav nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4719286570Smav 4720286570Smav nhdr->b_l1hdr.b_buf = buf; 4721286570Smav nhdr->b_l1hdr.b_datacnt = 1; 4722286570Smav nhdr->b_l1hdr.b_state = arc_anon; 4723286570Smav nhdr->b_l1hdr.b_arc_access = 0; 4724286763Smav nhdr->b_l1hdr.b_tmp_cdata = NULL; 4725168404Spjd nhdr->b_freeze_cksum = NULL; 4726286570Smav 4727286570Smav (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4728168404Spjd buf->b_hdr = nhdr; 4729219089Spjd mutex_exit(&buf->b_evict_lock); 4730168404Spjd atomic_add_64(&arc_anon->arcs_size, blksz); 4731168404Spjd } else { 4732219089Spjd mutex_exit(&buf->b_evict_lock); 4733286570Smav ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4734286763Smav /* protected by hash lock, or hdr is on arc_anon */ 4735286763Smav ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4736168404Spjd ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4737286570Smav arc_change_state(arc_anon, hdr, hash_lock); 4738286570Smav hdr->b_l1hdr.b_arc_access = 0; 4739286570Smav mutex_exit(hash_lock); 4740185029Spjd 4741219089Spjd buf_discard_identity(hdr); 4742168404Spjd arc_buf_thaw(buf); 4743168404Spjd } 4744168404Spjd buf->b_efunc = NULL; 4745168404Spjd buf->b_private = NULL; 4746168404Spjd} 4747168404Spjd 4748168404Spjdint 4749168404Spjdarc_released(arc_buf_t *buf) 4750168404Spjd{ 4751185029Spjd int released; 4752185029Spjd 4753219089Spjd mutex_enter(&buf->b_evict_lock); 4754286570Smav released = (buf->b_data != NULL && 4755286570Smav buf->b_hdr->b_l1hdr.b_state == arc_anon); 4756219089Spjd mutex_exit(&buf->b_evict_lock); 4757185029Spjd return (released); 4758168404Spjd} 4759168404Spjd 4760168404Spjd#ifdef ZFS_DEBUG 4761168404Spjdint 4762168404Spjdarc_referenced(arc_buf_t *buf) 4763168404Spjd{ 4764185029Spjd int referenced; 4765185029Spjd 4766219089Spjd mutex_enter(&buf->b_evict_lock); 4767286570Smav referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4768219089Spjd mutex_exit(&buf->b_evict_lock); 4769185029Spjd return (referenced); 4770168404Spjd} 4771168404Spjd#endif 4772168404Spjd 4773168404Spjdstatic void 4774168404Spjdarc_write_ready(zio_t *zio) 4775168404Spjd{ 4776168404Spjd arc_write_callback_t *callback = zio->io_private; 4777168404Spjd arc_buf_t *buf = callback->awcb_buf; 4778185029Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4779168404Spjd 4780286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 4781286570Smav ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4782286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4783185029Spjd callback->awcb_ready(zio, buf, callback->awcb_private); 4784185029Spjd 4785185029Spjd /* 4786185029Spjd * If the IO is already in progress, then this is a re-write 4787185029Spjd * attempt, so we need to thaw and re-compute the cksum. 4788185029Spjd * It is the responsibility of the callback to handle the 4789185029Spjd * accounting for any re-write attempt. 4790185029Spjd */ 4791185029Spjd if (HDR_IO_IN_PROGRESS(hdr)) { 4792286570Smav mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4793185029Spjd if (hdr->b_freeze_cksum != NULL) { 4794185029Spjd kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4795185029Spjd hdr->b_freeze_cksum = NULL; 4796185029Spjd } 4797286570Smav mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4798168404Spjd } 4799185029Spjd arc_cksum_compute(buf, B_FALSE); 4800275811Sdelphij hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4801168404Spjd} 4802168404Spjd 4803258632Savg/* 4804258632Savg * The SPA calls this callback for each physical write that happens on behalf 4805258632Savg * of a logical write. See the comment in dbuf_write_physdone() for details. 4806258632Savg */ 4807168404Spjdstatic void 4808258632Savgarc_write_physdone(zio_t *zio) 4809258632Savg{ 4810258632Savg arc_write_callback_t *cb = zio->io_private; 4811258632Savg if (cb->awcb_physdone != NULL) 4812258632Savg cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4813258632Savg} 4814258632Savg 4815258632Savgstatic void 4816168404Spjdarc_write_done(zio_t *zio) 4817168404Spjd{ 4818168404Spjd arc_write_callback_t *callback = zio->io_private; 4819168404Spjd arc_buf_t *buf = callback->awcb_buf; 4820168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4821168404Spjd 4822286570Smav ASSERT(hdr->b_l1hdr.b_acb == NULL); 4823168404Spjd 4824219089Spjd if (zio->io_error == 0) { 4825268075Sdelphij if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4826260150Sdelphij buf_discard_identity(hdr); 4827260150Sdelphij } else { 4828260150Sdelphij hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4829260150Sdelphij hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4830260150Sdelphij } 4831219089Spjd } else { 4832219089Spjd ASSERT(BUF_EMPTY(hdr)); 4833219089Spjd } 4834219089Spjd 4835168404Spjd /* 4836268075Sdelphij * If the block to be written was all-zero or compressed enough to be 4837268075Sdelphij * embedded in the BP, no write was performed so there will be no 4838268075Sdelphij * dva/birth/checksum. The buffer must therefore remain anonymous 4839268075Sdelphij * (and uncached). 4840168404Spjd */ 4841168404Spjd if (!BUF_EMPTY(hdr)) { 4842168404Spjd arc_buf_hdr_t *exists; 4843168404Spjd kmutex_t *hash_lock; 4844168404Spjd 4845219089Spjd ASSERT(zio->io_error == 0); 4846219089Spjd 4847168404Spjd arc_cksum_verify(buf); 4848168404Spjd 4849168404Spjd exists = buf_hash_insert(hdr, &hash_lock); 4850286570Smav if (exists != NULL) { 4851168404Spjd /* 4852168404Spjd * This can only happen if we overwrite for 4853168404Spjd * sync-to-convergence, because we remove 4854168404Spjd * buffers from the hash table when we arc_free(). 4855168404Spjd */ 4856219089Spjd if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4857219089Spjd if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4858219089Spjd panic("bad overwrite, hdr=%p exists=%p", 4859219089Spjd (void *)hdr, (void *)exists); 4860286570Smav ASSERT(refcount_is_zero( 4861286570Smav &exists->b_l1hdr.b_refcnt)); 4862219089Spjd arc_change_state(arc_anon, exists, hash_lock); 4863219089Spjd mutex_exit(hash_lock); 4864219089Spjd arc_hdr_destroy(exists); 4865219089Spjd exists = buf_hash_insert(hdr, &hash_lock); 4866219089Spjd ASSERT3P(exists, ==, NULL); 4867243524Smm } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4868243524Smm /* nopwrite */ 4869243524Smm ASSERT(zio->io_prop.zp_nopwrite); 4870243524Smm if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4871243524Smm panic("bad nopwrite, hdr=%p exists=%p", 4872243524Smm (void *)hdr, (void *)exists); 4873219089Spjd } else { 4874219089Spjd /* Dedup */ 4875286570Smav ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4876286570Smav ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4877219089Spjd ASSERT(BP_GET_DEDUP(zio->io_bp)); 4878219089Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4879219089Spjd } 4880168404Spjd } 4881275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4882185029Spjd /* if it's not anon, we are doing a scrub */ 4883286570Smav if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4884185029Spjd arc_access(hdr, hash_lock); 4885168404Spjd mutex_exit(hash_lock); 4886168404Spjd } else { 4887275811Sdelphij hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4888168404Spjd } 4889168404Spjd 4890286570Smav ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4891219089Spjd callback->awcb_done(zio, buf, callback->awcb_private); 4892168404Spjd 4893168404Spjd kmem_free(callback, sizeof (arc_write_callback_t)); 4894168404Spjd} 4895168404Spjd 4896168404Spjdzio_t * 4897219089Spjdarc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4898251478Sdelphij blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4899258632Savg const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4900258632Savg arc_done_func_t *done, void *private, zio_priority_t priority, 4901268123Sdelphij int zio_flags, const zbookmark_phys_t *zb) 4902168404Spjd{ 4903168404Spjd arc_buf_hdr_t *hdr = buf->b_hdr; 4904168404Spjd arc_write_callback_t *callback; 4905185029Spjd zio_t *zio; 4906168404Spjd 4907185029Spjd ASSERT(ready != NULL); 4908219089Spjd ASSERT(done != NULL); 4909168404Spjd ASSERT(!HDR_IO_ERROR(hdr)); 4910286570Smav ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4911286570Smav ASSERT(hdr->b_l1hdr.b_acb == NULL); 4912286570Smav ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4913185029Spjd if (l2arc) 4914275811Sdelphij hdr->b_flags |= ARC_FLAG_L2CACHE; 4915251478Sdelphij if (l2arc_compress) 4916275811Sdelphij hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4917168404Spjd callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4918168404Spjd callback->awcb_ready = ready; 4919258632Savg callback->awcb_physdone = physdone; 4920168404Spjd callback->awcb_done = done; 4921168404Spjd callback->awcb_private = private; 4922168404Spjd callback->awcb_buf = buf; 4923168404Spjd 4924219089Spjd zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4925258632Savg arc_write_ready, arc_write_physdone, arc_write_done, callback, 4926258632Savg priority, zio_flags, zb); 4927185029Spjd 4928168404Spjd return (zio); 4929168404Spjd} 4930168404Spjd 4931185029Spjdstatic int 4932258632Savgarc_memory_throttle(uint64_t reserve, uint64_t txg) 4933185029Spjd{ 4934185029Spjd#ifdef _KERNEL 4935272483Ssmh uint64_t available_memory = ptob(freemem); 4936185029Spjd static uint64_t page_load = 0; 4937185029Spjd static uint64_t last_txg = 0; 4938185029Spjd 4939272483Ssmh#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4940185029Spjd available_memory = 4941272483Ssmh MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4942185029Spjd#endif 4943258632Savg 4944272483Ssmh if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4945185029Spjd return (0); 4946185029Spjd 4947185029Spjd if (txg > last_txg) { 4948185029Spjd last_txg = txg; 4949185029Spjd page_load = 0; 4950185029Spjd } 4951185029Spjd /* 4952185029Spjd * If we are in pageout, we know that memory is already tight, 4953185029Spjd * the arc is already going to be evicting, so we just want to 4954185029Spjd * continue to let page writes occur as quickly as possible. 4955185029Spjd */ 4956185029Spjd if (curproc == pageproc) { 4957272483Ssmh if (page_load > MAX(ptob(minfree), available_memory) / 4) 4958249195Smm return (SET_ERROR(ERESTART)); 4959185029Spjd /* Note: reserve is inflated, so we deflate */ 4960185029Spjd page_load += reserve / 8; 4961185029Spjd return (0); 4962185029Spjd } else if (page_load > 0 && arc_reclaim_needed()) { 4963185029Spjd /* memory is low, delay before restarting */ 4964185029Spjd ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4965249195Smm return (SET_ERROR(EAGAIN)); 4966185029Spjd } 4967185029Spjd page_load = 0; 4968185029Spjd#endif 4969185029Spjd return (0); 4970185029Spjd} 4971185029Spjd 4972168404Spjdvoid 4973185029Spjdarc_tempreserve_clear(uint64_t reserve) 4974168404Spjd{ 4975185029Spjd atomic_add_64(&arc_tempreserve, -reserve); 4976168404Spjd ASSERT((int64_t)arc_tempreserve >= 0); 4977168404Spjd} 4978168404Spjd 4979168404Spjdint 4980185029Spjdarc_tempreserve_space(uint64_t reserve, uint64_t txg) 4981168404Spjd{ 4982185029Spjd int error; 4983209962Smm uint64_t anon_size; 4984185029Spjd 4985272483Ssmh if (reserve > arc_c/4 && !arc_no_grow) { 4986185029Spjd arc_c = MIN(arc_c_max, reserve * 4); 4987272483Ssmh DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4988272483Ssmh } 4989185029Spjd if (reserve > arc_c) 4990249195Smm return (SET_ERROR(ENOMEM)); 4991168404Spjd 4992168404Spjd /* 4993209962Smm * Don't count loaned bufs as in flight dirty data to prevent long 4994209962Smm * network delays from blocking transactions that are ready to be 4995209962Smm * assigned to a txg. 4996209962Smm */ 4997209962Smm anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4998209962Smm 4999209962Smm /* 5000185029Spjd * Writes will, almost always, require additional memory allocations 5001251631Sdelphij * in order to compress/encrypt/etc the data. We therefore need to 5002185029Spjd * make sure that there is sufficient available memory for this. 5003185029Spjd */ 5004258632Savg error = arc_memory_throttle(reserve, txg); 5005258632Savg if (error != 0) 5006185029Spjd return (error); 5007185029Spjd 5008185029Spjd /* 5009168404Spjd * Throttle writes when the amount of dirty data in the cache 5010168404Spjd * gets too large. We try to keep the cache less than half full 5011168404Spjd * of dirty blocks so that our sync times don't grow too large. 5012168404Spjd * Note: if two requests come in concurrently, we might let them 5013168404Spjd * both succeed, when one of them should fail. Not a huge deal. 5014168404Spjd */ 5015209962Smm 5016209962Smm if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5017209962Smm anon_size > arc_c / 4) { 5018185029Spjd dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5019185029Spjd "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5020185029Spjd arc_tempreserve>>10, 5021185029Spjd arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5022185029Spjd arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5023185029Spjd reserve>>10, arc_c>>10); 5024249195Smm return (SET_ERROR(ERESTART)); 5025168404Spjd } 5026185029Spjd atomic_add_64(&arc_tempreserve, reserve); 5027168404Spjd return (0); 5028168404Spjd} 5029168404Spjd 5030286626Smavstatic void 5031286626Smavarc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5032286626Smav kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5033286626Smav{ 5034286626Smav size->value.ui64 = state->arcs_size; 5035286626Smav evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5036286626Smav evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5037286626Smav} 5038286626Smav 5039286626Smavstatic int 5040286626Smavarc_kstat_update(kstat_t *ksp, int rw) 5041286626Smav{ 5042286626Smav arc_stats_t *as = ksp->ks_data; 5043286626Smav 5044286626Smav if (rw == KSTAT_WRITE) { 5045286626Smav return (EACCES); 5046286626Smav } else { 5047286626Smav arc_kstat_update_state(arc_anon, 5048286626Smav &as->arcstat_anon_size, 5049286626Smav &as->arcstat_anon_evictable_data, 5050286626Smav &as->arcstat_anon_evictable_metadata); 5051286626Smav arc_kstat_update_state(arc_mru, 5052286626Smav &as->arcstat_mru_size, 5053286626Smav &as->arcstat_mru_evictable_data, 5054286626Smav &as->arcstat_mru_evictable_metadata); 5055286626Smav arc_kstat_update_state(arc_mru_ghost, 5056286626Smav &as->arcstat_mru_ghost_size, 5057286626Smav &as->arcstat_mru_ghost_evictable_data, 5058286626Smav &as->arcstat_mru_ghost_evictable_metadata); 5059286626Smav arc_kstat_update_state(arc_mfu, 5060286626Smav &as->arcstat_mfu_size, 5061286626Smav &as->arcstat_mfu_evictable_data, 5062286626Smav &as->arcstat_mfu_evictable_metadata); 5063286626Smav arc_kstat_update_state(arc_mfu_ghost, 5064286626Smav &as->arcstat_mfu_ghost_size, 5065286626Smav &as->arcstat_mfu_ghost_evictable_data, 5066286626Smav &as->arcstat_mfu_ghost_evictable_metadata); 5067286626Smav } 5068286626Smav 5069286626Smav return (0); 5070286626Smav} 5071286626Smav 5072286763Smav/* 5073286763Smav * This function *must* return indices evenly distributed between all 5074286763Smav * sublists of the multilist. This is needed due to how the ARC eviction 5075286763Smav * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5076286763Smav * distributed between all sublists and uses this assumption when 5077286763Smav * deciding which sublist to evict from and how much to evict from it. 5078286763Smav */ 5079286763Smavunsigned int 5080286763Smavarc_state_multilist_index_func(multilist_t *ml, void *obj) 5081286763Smav{ 5082286763Smav arc_buf_hdr_t *hdr = obj; 5083286763Smav 5084286763Smav /* 5085286763Smav * We rely on b_dva to generate evenly distributed index 5086286763Smav * numbers using buf_hash below. So, as an added precaution, 5087286763Smav * let's make sure we never add empty buffers to the arc lists. 5088286763Smav */ 5089286763Smav ASSERT(!BUF_EMPTY(hdr)); 5090286763Smav 5091286763Smav /* 5092286763Smav * The assumption here, is the hash value for a given 5093286763Smav * arc_buf_hdr_t will remain constant throughout it's lifetime 5094286763Smav * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5095286763Smav * Thus, we don't need to store the header's sublist index 5096286763Smav * on insertion, as this index can be recalculated on removal. 5097286763Smav * 5098286763Smav * Also, the low order bits of the hash value are thought to be 5099286763Smav * distributed evenly. Otherwise, in the case that the multilist 5100286763Smav * has a power of two number of sublists, each sublists' usage 5101286763Smav * would not be evenly distributed. 5102286763Smav */ 5103286763Smav return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5104286763Smav multilist_get_num_sublists(ml)); 5105286763Smav} 5106286763Smav 5107168404Spjd#ifdef _KERNEL 5108168566Spjdstatic eventhandler_tag arc_event_lowmem = NULL; 5109168404Spjd 5110168404Spjdstatic void 5111168566Spjdarc_lowmem(void *arg __unused, int howto __unused) 5112168404Spjd{ 5113168404Spjd 5114286763Smav mutex_enter(&arc_reclaim_lock); 5115286625Smav /* XXX: Memory deficit should be passed as argument. */ 5116286625Smav needfree = btoc(arc_c >> arc_shrink_shift); 5117272483Ssmh DTRACE_PROBE(arc__needfree); 5118286763Smav cv_signal(&arc_reclaim_thread_cv); 5119241773Savg 5120241773Savg /* 5121241773Savg * It is unsafe to block here in arbitrary threads, because we can come 5122241773Savg * here from ARC itself and may hold ARC locks and thus risk a deadlock 5123241773Savg * with ARC reclaim thread. 5124241773Savg */ 5125286623Smav if (curproc == pageproc) 5126286763Smav (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5127286763Smav mutex_exit(&arc_reclaim_lock); 5128168404Spjd} 5129168404Spjd#endif 5130168404Spjd 5131168404Spjdvoid 5132168404Spjdarc_init(void) 5133168404Spjd{ 5134219089Spjd int i, prefetch_tunable_set = 0; 5135205231Skmacy 5136286763Smav mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5137286763Smav cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5138286763Smav cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5139168404Spjd 5140286763Smav mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5141286763Smav cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5142286763Smav 5143168404Spjd /* Convert seconds to clock ticks */ 5144168404Spjd arc_min_prefetch_lifespan = 1 * hz; 5145168404Spjd 5146168404Spjd /* Start out with 1/8 of all memory */ 5147168566Spjd arc_c = kmem_size() / 8; 5148219089Spjd 5149277300Ssmh#ifdef illumos 5150192360Skmacy#ifdef _KERNEL 5151192360Skmacy /* 5152192360Skmacy * On architectures where the physical memory can be larger 5153192360Skmacy * than the addressable space (intel in 32-bit mode), we may 5154192360Skmacy * need to limit the cache to 1/8 of VM size. 5155192360Skmacy */ 5156192360Skmacy arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5157192360Skmacy#endif 5158277300Ssmh#endif /* illumos */ 5159168566Spjd /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5160280822Smav arc_c_min = MAX(arc_c / 4, 16 << 20); 5161168566Spjd /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5162280822Smav if (arc_c * 8 >= 1 << 30) 5163280822Smav arc_c_max = (arc_c * 8) - (1 << 30); 5164168404Spjd else 5165168404Spjd arc_c_max = arc_c_min; 5166175633Spjd arc_c_max = MAX(arc_c * 5, arc_c_max); 5167219089Spjd 5168168481Spjd#ifdef _KERNEL 5169168404Spjd /* 5170168404Spjd * Allow the tunables to override our calculations if they are 5171168566Spjd * reasonable (ie. over 16MB) 5172168404Spjd */ 5173280822Smav if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5174168404Spjd arc_c_max = zfs_arc_max; 5175280822Smav if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5176168404Spjd arc_c_min = zfs_arc_min; 5177168481Spjd#endif 5178219089Spjd 5179168404Spjd arc_c = arc_c_max; 5180168404Spjd arc_p = (arc_c >> 1); 5181168404Spjd 5182185029Spjd /* limit meta-data to 1/4 of the arc capacity */ 5183185029Spjd arc_meta_limit = arc_c_max / 4; 5184185029Spjd 5185185029Spjd /* Allow the tunable to override if it is reasonable */ 5186185029Spjd if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5187185029Spjd arc_meta_limit = zfs_arc_meta_limit; 5188185029Spjd 5189185029Spjd if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5190185029Spjd arc_c_min = arc_meta_limit / 2; 5191185029Spjd 5192275780Sdelphij if (zfs_arc_meta_min > 0) { 5193275780Sdelphij arc_meta_min = zfs_arc_meta_min; 5194275780Sdelphij } else { 5195275780Sdelphij arc_meta_min = arc_c_min / 2; 5196275780Sdelphij } 5197275780Sdelphij 5198208373Smm if (zfs_arc_grow_retry > 0) 5199208373Smm arc_grow_retry = zfs_arc_grow_retry; 5200208373Smm 5201208373Smm if (zfs_arc_shrink_shift > 0) 5202208373Smm arc_shrink_shift = zfs_arc_shrink_shift; 5203208373Smm 5204286625Smav /* 5205286625Smav * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5206286625Smav */ 5207286625Smav if (arc_no_grow_shift >= arc_shrink_shift) 5208286625Smav arc_no_grow_shift = arc_shrink_shift - 1; 5209286625Smav 5210208373Smm if (zfs_arc_p_min_shift > 0) 5211208373Smm arc_p_min_shift = zfs_arc_p_min_shift; 5212208373Smm 5213286763Smav if (zfs_arc_num_sublists_per_state < 1) 5214286763Smav zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5215286763Smav 5216168404Spjd /* if kmem_flags are set, lets try to use less memory */ 5217168404Spjd if (kmem_debugging()) 5218168404Spjd arc_c = arc_c / 2; 5219168404Spjd if (arc_c < arc_c_min) 5220168404Spjd arc_c = arc_c_min; 5221168404Spjd 5222168473Spjd zfs_arc_min = arc_c_min; 5223168473Spjd zfs_arc_max = arc_c_max; 5224168473Spjd 5225168404Spjd arc_anon = &ARC_anon; 5226168404Spjd arc_mru = &ARC_mru; 5227168404Spjd arc_mru_ghost = &ARC_mru_ghost; 5228168404Spjd arc_mfu = &ARC_mfu; 5229168404Spjd arc_mfu_ghost = &ARC_mfu_ghost; 5230185029Spjd arc_l2c_only = &ARC_l2c_only; 5231168404Spjd arc_size = 0; 5232168404Spjd 5233286763Smav multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5234286762Smav sizeof (arc_buf_hdr_t), 5235286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5236286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5237286763Smav multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5238286762Smav sizeof (arc_buf_hdr_t), 5239286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5240286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5241286763Smav multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5242286762Smav sizeof (arc_buf_hdr_t), 5243286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5244286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5245286763Smav multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5246286762Smav sizeof (arc_buf_hdr_t), 5247286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5248286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5249286763Smav multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5250286762Smav sizeof (arc_buf_hdr_t), 5251286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5252286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5253286763Smav multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5254286762Smav sizeof (arc_buf_hdr_t), 5255286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5256286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5257286763Smav multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5258286762Smav sizeof (arc_buf_hdr_t), 5259286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5260286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5261286763Smav multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5262286762Smav sizeof (arc_buf_hdr_t), 5263286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5264286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5265286763Smav multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5266286762Smav sizeof (arc_buf_hdr_t), 5267286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5268286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5269286763Smav multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5270286762Smav sizeof (arc_buf_hdr_t), 5271286763Smav offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5272286763Smav zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5273168404Spjd 5274168404Spjd buf_init(); 5275168404Spjd 5276286763Smav arc_reclaim_thread_exit = FALSE; 5277286763Smav arc_user_evicts_thread_exit = FALSE; 5278168404Spjd arc_eviction_list = NULL; 5279168404Spjd bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5280168404Spjd 5281168404Spjd arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5282168404Spjd sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5283168404Spjd 5284168404Spjd if (arc_ksp != NULL) { 5285168404Spjd arc_ksp->ks_data = &arc_stats; 5286286574Smav arc_ksp->ks_update = arc_kstat_update; 5287168404Spjd kstat_install(arc_ksp); 5288168404Spjd } 5289168404Spjd 5290168404Spjd (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5291168404Spjd TS_RUN, minclsyspri); 5292168404Spjd 5293168404Spjd#ifdef _KERNEL 5294168566Spjd arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5295168404Spjd EVENTHANDLER_PRI_FIRST); 5296168404Spjd#endif 5297168404Spjd 5298286763Smav (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5299286763Smav TS_RUN, minclsyspri); 5300286763Smav 5301168404Spjd arc_dead = FALSE; 5302185029Spjd arc_warm = B_FALSE; 5303168566Spjd 5304258632Savg /* 5305258632Savg * Calculate maximum amount of dirty data per pool. 5306258632Savg * 5307258632Savg * If it has been set by /etc/system, take that. 5308258632Savg * Otherwise, use a percentage of physical memory defined by 5309258632Savg * zfs_dirty_data_max_percent (default 10%) with a cap at 5310258632Savg * zfs_dirty_data_max_max (default 4GB). 5311258632Savg */ 5312258632Savg if (zfs_dirty_data_max == 0) { 5313258632Savg zfs_dirty_data_max = ptob(physmem) * 5314258632Savg zfs_dirty_data_max_percent / 100; 5315258632Savg zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5316258632Savg zfs_dirty_data_max_max); 5317258632Savg } 5318185029Spjd 5319168566Spjd#ifdef _KERNEL 5320194043Skmacy if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5321193953Skmacy prefetch_tunable_set = 1; 5322206796Spjd 5323193878Skmacy#ifdef __i386__ 5324193953Skmacy if (prefetch_tunable_set == 0) { 5325196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5326196863Strasz "-- to enable,\n"); 5327196863Strasz printf(" add \"vfs.zfs.prefetch_disable=0\" " 5328196863Strasz "to /boot/loader.conf.\n"); 5329219089Spjd zfs_prefetch_disable = 1; 5330193878Skmacy } 5331206796Spjd#else 5332193878Skmacy if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5333193953Skmacy prefetch_tunable_set == 0) { 5334196863Strasz printf("ZFS NOTICE: Prefetch is disabled by default if less " 5335196941Strasz "than 4GB of RAM is present;\n" 5336196863Strasz " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5337196863Strasz "to /boot/loader.conf.\n"); 5338219089Spjd zfs_prefetch_disable = 1; 5339193878Skmacy } 5340206796Spjd#endif 5341175633Spjd /* Warn about ZFS memory and address space requirements. */ 5342168696Spjd if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5343168987Sbmah printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5344168987Sbmah "expect unstable behavior.\n"); 5345175633Spjd } 5346175633Spjd if (kmem_size() < 512 * (1 << 20)) { 5347173419Spjd printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5348168987Sbmah "expect unstable behavior.\n"); 5349185029Spjd printf(" Consider tuning vm.kmem_size and " 5350173419Spjd "vm.kmem_size_max\n"); 5351185029Spjd printf(" in /boot/loader.conf.\n"); 5352168566Spjd } 5353168566Spjd#endif 5354168404Spjd} 5355168404Spjd 5356168404Spjdvoid 5357168404Spjdarc_fini(void) 5358168404Spjd{ 5359286763Smav mutex_enter(&arc_reclaim_lock); 5360286763Smav arc_reclaim_thread_exit = TRUE; 5361286763Smav /* 5362286763Smav * The reclaim thread will set arc_reclaim_thread_exit back to 5363286763Smav * FALSE when it is finished exiting; we're waiting for that. 5364286763Smav */ 5365286763Smav while (arc_reclaim_thread_exit) { 5366286763Smav cv_signal(&arc_reclaim_thread_cv); 5367286763Smav cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5368286763Smav } 5369286763Smav mutex_exit(&arc_reclaim_lock); 5370168404Spjd 5371286763Smav mutex_enter(&arc_user_evicts_lock); 5372286763Smav arc_user_evicts_thread_exit = TRUE; 5373286763Smav /* 5374286763Smav * The user evicts thread will set arc_user_evicts_thread_exit 5375286763Smav * to FALSE when it is finished exiting; we're waiting for that. 5376286763Smav */ 5377286763Smav while (arc_user_evicts_thread_exit) { 5378286763Smav cv_signal(&arc_user_evicts_cv); 5379286763Smav cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5380286763Smav } 5381286763Smav mutex_exit(&arc_user_evicts_lock); 5382168404Spjd 5383286763Smav /* Use TRUE to ensure *all* buffers are evicted */ 5384286763Smav arc_flush(NULL, TRUE); 5385286763Smav 5386168404Spjd arc_dead = TRUE; 5387168404Spjd 5388168404Spjd if (arc_ksp != NULL) { 5389168404Spjd kstat_delete(arc_ksp); 5390168404Spjd arc_ksp = NULL; 5391168404Spjd } 5392168404Spjd 5393286763Smav mutex_destroy(&arc_reclaim_lock); 5394286763Smav cv_destroy(&arc_reclaim_thread_cv); 5395286763Smav cv_destroy(&arc_reclaim_waiters_cv); 5396168404Spjd 5397286763Smav mutex_destroy(&arc_user_evicts_lock); 5398286763Smav cv_destroy(&arc_user_evicts_cv); 5399168404Spjd 5400286763Smav multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5401286763Smav multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5402286763Smav multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5403286763Smav multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5404286763Smav multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5405286763Smav multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5406286763Smav multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5407286763Smav multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5408206796Spjd 5409168404Spjd buf_fini(); 5410168404Spjd 5411286570Smav ASSERT0(arc_loaned_bytes); 5412209962Smm 5413168404Spjd#ifdef _KERNEL 5414168566Spjd if (arc_event_lowmem != NULL) 5415168566Spjd EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5416168404Spjd#endif 5417168404Spjd} 5418185029Spjd 5419185029Spjd/* 5420185029Spjd * Level 2 ARC 5421185029Spjd * 5422185029Spjd * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5423185029Spjd * It uses dedicated storage devices to hold cached data, which are populated 5424185029Spjd * using large infrequent writes. The main role of this cache is to boost 5425185029Spjd * the performance of random read workloads. The intended L2ARC devices 5426185029Spjd * include short-stroked disks, solid state disks, and other media with 5427185029Spjd * substantially faster read latency than disk. 5428185029Spjd * 5429185029Spjd * +-----------------------+ 5430185029Spjd * | ARC | 5431185029Spjd * +-----------------------+ 5432185029Spjd * | ^ ^ 5433185029Spjd * | | | 5434185029Spjd * l2arc_feed_thread() arc_read() 5435185029Spjd * | | | 5436185029Spjd * | l2arc read | 5437185029Spjd * V | | 5438185029Spjd * +---------------+ | 5439185029Spjd * | L2ARC | | 5440185029Spjd * +---------------+ | 5441185029Spjd * | ^ | 5442185029Spjd * l2arc_write() | | 5443185029Spjd * | | | 5444185029Spjd * V | | 5445185029Spjd * +-------+ +-------+ 5446185029Spjd * | vdev | | vdev | 5447185029Spjd * | cache | | cache | 5448185029Spjd * +-------+ +-------+ 5449185029Spjd * +=========+ .-----. 5450185029Spjd * : L2ARC : |-_____-| 5451185029Spjd * : devices : | Disks | 5452185029Spjd * +=========+ `-_____-' 5453185029Spjd * 5454185029Spjd * Read requests are satisfied from the following sources, in order: 5455185029Spjd * 5456185029Spjd * 1) ARC 5457185029Spjd * 2) vdev cache of L2ARC devices 5458185029Spjd * 3) L2ARC devices 5459185029Spjd * 4) vdev cache of disks 5460185029Spjd * 5) disks 5461185029Spjd * 5462185029Spjd * Some L2ARC device types exhibit extremely slow write performance. 5463185029Spjd * To accommodate for this there are some significant differences between 5464185029Spjd * the L2ARC and traditional cache design: 5465185029Spjd * 5466185029Spjd * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5467185029Spjd * the ARC behave as usual, freeing buffers and placing headers on ghost 5468185029Spjd * lists. The ARC does not send buffers to the L2ARC during eviction as 5469185029Spjd * this would add inflated write latencies for all ARC memory pressure. 5470185029Spjd * 5471185029Spjd * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5472185029Spjd * It does this by periodically scanning buffers from the eviction-end of 5473185029Spjd * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5474251478Sdelphij * not already there. It scans until a headroom of buffers is satisfied, 5475251478Sdelphij * which itself is a buffer for ARC eviction. If a compressible buffer is 5476251478Sdelphij * found during scanning and selected for writing to an L2ARC device, we 5477251478Sdelphij * temporarily boost scanning headroom during the next scan cycle to make 5478251478Sdelphij * sure we adapt to compression effects (which might significantly reduce 5479251478Sdelphij * the data volume we write to L2ARC). The thread that does this is 5480185029Spjd * l2arc_feed_thread(), illustrated below; example sizes are included to 5481185029Spjd * provide a better sense of ratio than this diagram: 5482185029Spjd * 5483185029Spjd * head --> tail 5484185029Spjd * +---------------------+----------+ 5485185029Spjd * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5486185029Spjd * +---------------------+----------+ | o L2ARC eligible 5487185029Spjd * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5488185029Spjd * +---------------------+----------+ | 5489185029Spjd * 15.9 Gbytes ^ 32 Mbytes | 5490185029Spjd * headroom | 5491185029Spjd * l2arc_feed_thread() 5492185029Spjd * | 5493185029Spjd * l2arc write hand <--[oooo]--' 5494185029Spjd * | 8 Mbyte 5495185029Spjd * | write max 5496185029Spjd * V 5497185029Spjd * +==============================+ 5498185029Spjd * L2ARC dev |####|#|###|###| |####| ... | 5499185029Spjd * +==============================+ 5500185029Spjd * 32 Gbytes 5501185029Spjd * 5502185029Spjd * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5503185029Spjd * evicted, then the L2ARC has cached a buffer much sooner than it probably 5504185029Spjd * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5505185029Spjd * safe to say that this is an uncommon case, since buffers at the end of 5506185029Spjd * the ARC lists have moved there due to inactivity. 5507185029Spjd * 5508185029Spjd * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5509185029Spjd * then the L2ARC simply misses copying some buffers. This serves as a 5510185029Spjd * pressure valve to prevent heavy read workloads from both stalling the ARC 5511185029Spjd * with waits and clogging the L2ARC with writes. This also helps prevent 5512185029Spjd * the potential for the L2ARC to churn if it attempts to cache content too 5513185029Spjd * quickly, such as during backups of the entire pool. 5514185029Spjd * 5515185029Spjd * 5. After system boot and before the ARC has filled main memory, there are 5516185029Spjd * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5517185029Spjd * lists can remain mostly static. Instead of searching from tail of these 5518185029Spjd * lists as pictured, the l2arc_feed_thread() will search from the list heads 5519185029Spjd * for eligible buffers, greatly increasing its chance of finding them. 5520185029Spjd * 5521185029Spjd * The L2ARC device write speed is also boosted during this time so that 5522185029Spjd * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5523185029Spjd * there are no L2ARC reads, and no fear of degrading read performance 5524185029Spjd * through increased writes. 5525185029Spjd * 5526185029Spjd * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5527185029Spjd * the vdev queue can aggregate them into larger and fewer writes. Each 5528185029Spjd * device is written to in a rotor fashion, sweeping writes through 5529185029Spjd * available space then repeating. 5530185029Spjd * 5531185029Spjd * 7. The L2ARC does not store dirty content. It never needs to flush 5532185029Spjd * write buffers back to disk based storage. 5533185029Spjd * 5534185029Spjd * 8. If an ARC buffer is written (and dirtied) which also exists in the 5535185029Spjd * L2ARC, the now stale L2ARC buffer is immediately dropped. 5536185029Spjd * 5537185029Spjd * The performance of the L2ARC can be tweaked by a number of tunables, which 5538185029Spjd * may be necessary for different workloads: 5539185029Spjd * 5540185029Spjd * l2arc_write_max max write bytes per interval 5541185029Spjd * l2arc_write_boost extra write bytes during device warmup 5542185029Spjd * l2arc_noprefetch skip caching prefetched buffers 5543185029Spjd * l2arc_headroom number of max device writes to precache 5544251478Sdelphij * l2arc_headroom_boost when we find compressed buffers during ARC 5545251478Sdelphij * scanning, we multiply headroom by this 5546251478Sdelphij * percentage factor for the next scan cycle, 5547251478Sdelphij * since more compressed buffers are likely to 5548251478Sdelphij * be present 5549185029Spjd * l2arc_feed_secs seconds between L2ARC writing 5550185029Spjd * 5551185029Spjd * Tunables may be removed or added as future performance improvements are 5552185029Spjd * integrated, and also may become zpool properties. 5553208373Smm * 5554208373Smm * There are three key functions that control how the L2ARC warms up: 5555208373Smm * 5556208373Smm * l2arc_write_eligible() check if a buffer is eligible to cache 5557208373Smm * l2arc_write_size() calculate how much to write 5558208373Smm * l2arc_write_interval() calculate sleep delay between writes 5559208373Smm * 5560208373Smm * These three functions determine what to write, how much, and how quickly 5561208373Smm * to send writes. 5562185029Spjd */ 5563185029Spjd 5564208373Smmstatic boolean_t 5565275811Sdelphijl2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5566208373Smm{ 5567208373Smm /* 5568208373Smm * A buffer is *not* eligible for the L2ARC if it: 5569208373Smm * 1. belongs to a different spa. 5570208373Smm * 2. is already cached on the L2ARC. 5571208373Smm * 3. has an I/O in progress (it may be an incomplete read). 5572208373Smm * 4. is flagged not eligible (zfs property). 5573208373Smm */ 5574275811Sdelphij if (hdr->b_spa != spa_guid) { 5575208373Smm ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5576208373Smm return (B_FALSE); 5577208373Smm } 5578286570Smav if (HDR_HAS_L2HDR(hdr)) { 5579208373Smm ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5580208373Smm return (B_FALSE); 5581208373Smm } 5582275811Sdelphij if (HDR_IO_IN_PROGRESS(hdr)) { 5583208373Smm ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5584208373Smm return (B_FALSE); 5585208373Smm } 5586275811Sdelphij if (!HDR_L2CACHE(hdr)) { 5587208373Smm ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5588208373Smm return (B_FALSE); 5589208373Smm } 5590208373Smm 5591208373Smm return (B_TRUE); 5592208373Smm} 5593208373Smm 5594208373Smmstatic uint64_t 5595251478Sdelphijl2arc_write_size(void) 5596208373Smm{ 5597208373Smm uint64_t size; 5598208373Smm 5599251478Sdelphij /* 5600251478Sdelphij * Make sure our globals have meaningful values in case the user 5601251478Sdelphij * altered them. 5602251478Sdelphij */ 5603251478Sdelphij size = l2arc_write_max; 5604251478Sdelphij if (size == 0) { 5605251478Sdelphij cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5606251478Sdelphij "be greater than zero, resetting it to the default (%d)", 5607251478Sdelphij L2ARC_WRITE_SIZE); 5608251478Sdelphij size = l2arc_write_max = L2ARC_WRITE_SIZE; 5609251478Sdelphij } 5610208373Smm 5611208373Smm if (arc_warm == B_FALSE) 5612251478Sdelphij size += l2arc_write_boost; 5613208373Smm 5614208373Smm return (size); 5615208373Smm 5616208373Smm} 5617208373Smm 5618208373Smmstatic clock_t 5619208373Smml2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5620208373Smm{ 5621219089Spjd clock_t interval, next, now; 5622208373Smm 5623208373Smm /* 5624208373Smm * If the ARC lists are busy, increase our write rate; if the 5625208373Smm * lists are stale, idle back. This is achieved by checking 5626208373Smm * how much we previously wrote - if it was more than half of 5627208373Smm * what we wanted, schedule the next write much sooner. 5628208373Smm */ 5629208373Smm if (l2arc_feed_again && wrote > (wanted / 2)) 5630208373Smm interval = (hz * l2arc_feed_min_ms) / 1000; 5631208373Smm else 5632208373Smm interval = hz * l2arc_feed_secs; 5633208373Smm 5634219089Spjd now = ddi_get_lbolt(); 5635219089Spjd next = MAX(now, MIN(now + interval, began + interval)); 5636208373Smm 5637208373Smm return (next); 5638208373Smm} 5639208373Smm 5640185029Spjd/* 5641185029Spjd * Cycle through L2ARC devices. This is how L2ARC load balances. 5642185029Spjd * If a device is returned, this also returns holding the spa config lock. 5643185029Spjd */ 5644185029Spjdstatic l2arc_dev_t * 5645185029Spjdl2arc_dev_get_next(void) 5646185029Spjd{ 5647185029Spjd l2arc_dev_t *first, *next = NULL; 5648185029Spjd 5649185029Spjd /* 5650185029Spjd * Lock out the removal of spas (spa_namespace_lock), then removal 5651185029Spjd * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5652185029Spjd * both locks will be dropped and a spa config lock held instead. 5653185029Spjd */ 5654185029Spjd mutex_enter(&spa_namespace_lock); 5655185029Spjd mutex_enter(&l2arc_dev_mtx); 5656185029Spjd 5657185029Spjd /* if there are no vdevs, there is nothing to do */ 5658185029Spjd if (l2arc_ndev == 0) 5659185029Spjd goto out; 5660185029Spjd 5661185029Spjd first = NULL; 5662185029Spjd next = l2arc_dev_last; 5663185029Spjd do { 5664185029Spjd /* loop around the list looking for a non-faulted vdev */ 5665185029Spjd if (next == NULL) { 5666185029Spjd next = list_head(l2arc_dev_list); 5667185029Spjd } else { 5668185029Spjd next = list_next(l2arc_dev_list, next); 5669185029Spjd if (next == NULL) 5670185029Spjd next = list_head(l2arc_dev_list); 5671185029Spjd } 5672185029Spjd 5673185029Spjd /* if we have come back to the start, bail out */ 5674185029Spjd if (first == NULL) 5675185029Spjd first = next; 5676185029Spjd else if (next == first) 5677185029Spjd break; 5678185029Spjd 5679185029Spjd } while (vdev_is_dead(next->l2ad_vdev)); 5680185029Spjd 5681185029Spjd /* if we were unable to find any usable vdevs, return NULL */ 5682185029Spjd if (vdev_is_dead(next->l2ad_vdev)) 5683185029Spjd next = NULL; 5684185029Spjd 5685185029Spjd l2arc_dev_last = next; 5686185029Spjd 5687185029Spjdout: 5688185029Spjd mutex_exit(&l2arc_dev_mtx); 5689185029Spjd 5690185029Spjd /* 5691185029Spjd * Grab the config lock to prevent the 'next' device from being 5692185029Spjd * removed while we are writing to it. 5693185029Spjd */ 5694185029Spjd if (next != NULL) 5695185029Spjd spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5696185029Spjd mutex_exit(&spa_namespace_lock); 5697185029Spjd 5698185029Spjd return (next); 5699185029Spjd} 5700185029Spjd 5701185029Spjd/* 5702185029Spjd * Free buffers that were tagged for destruction. 5703185029Spjd */ 5704185029Spjdstatic void 5705185029Spjdl2arc_do_free_on_write() 5706185029Spjd{ 5707185029Spjd list_t *buflist; 5708185029Spjd l2arc_data_free_t *df, *df_prev; 5709185029Spjd 5710185029Spjd mutex_enter(&l2arc_free_on_write_mtx); 5711185029Spjd buflist = l2arc_free_on_write; 5712185029Spjd 5713185029Spjd for (df = list_tail(buflist); df; df = df_prev) { 5714185029Spjd df_prev = list_prev(buflist, df); 5715185029Spjd ASSERT(df->l2df_data != NULL); 5716185029Spjd ASSERT(df->l2df_func != NULL); 5717185029Spjd df->l2df_func(df->l2df_data, df->l2df_size); 5718185029Spjd list_remove(buflist, df); 5719185029Spjd kmem_free(df, sizeof (l2arc_data_free_t)); 5720185029Spjd } 5721185029Spjd 5722185029Spjd mutex_exit(&l2arc_free_on_write_mtx); 5723185029Spjd} 5724185029Spjd 5725185029Spjd/* 5726185029Spjd * A write to a cache device has completed. Update all headers to allow 5727185029Spjd * reads from these buffers to begin. 5728185029Spjd */ 5729185029Spjdstatic void 5730185029Spjdl2arc_write_done(zio_t *zio) 5731185029Spjd{ 5732185029Spjd l2arc_write_callback_t *cb; 5733185029Spjd l2arc_dev_t *dev; 5734185029Spjd list_t *buflist; 5735275811Sdelphij arc_buf_hdr_t *head, *hdr, *hdr_prev; 5736185029Spjd kmutex_t *hash_lock; 5737268085Sdelphij int64_t bytes_dropped = 0; 5738185029Spjd 5739185029Spjd cb = zio->io_private; 5740185029Spjd ASSERT(cb != NULL); 5741185029Spjd dev = cb->l2wcb_dev; 5742185029Spjd ASSERT(dev != NULL); 5743185029Spjd head = cb->l2wcb_head; 5744185029Spjd ASSERT(head != NULL); 5745286570Smav buflist = &dev->l2ad_buflist; 5746185029Spjd ASSERT(buflist != NULL); 5747185029Spjd DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5748185029Spjd l2arc_write_callback_t *, cb); 5749185029Spjd 5750185029Spjd if (zio->io_error != 0) 5751185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_error); 5752185029Spjd 5753185029Spjd /* 5754185029Spjd * All writes completed, or an error was hit. 5755185029Spjd */ 5756286763Smavtop: 5757286763Smav mutex_enter(&dev->l2ad_mtx); 5758275811Sdelphij for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5759275811Sdelphij hdr_prev = list_prev(buflist, hdr); 5760185029Spjd 5761275811Sdelphij hash_lock = HDR_LOCK(hdr); 5762286763Smav 5763286763Smav /* 5764286763Smav * We cannot use mutex_enter or else we can deadlock 5765286763Smav * with l2arc_write_buffers (due to swapping the order 5766286763Smav * the hash lock and l2ad_mtx are taken). 5767286763Smav */ 5768185029Spjd if (!mutex_tryenter(hash_lock)) { 5769185029Spjd /* 5770286763Smav * Missed the hash lock. We must retry so we 5771286763Smav * don't leave the ARC_FLAG_L2_WRITING bit set. 5772185029Spjd */ 5773286763Smav ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5774286763Smav 5775286763Smav /* 5776286763Smav * We don't want to rescan the headers we've 5777286763Smav * already marked as having been written out, so 5778286763Smav * we reinsert the head node so we can pick up 5779286763Smav * where we left off. 5780286763Smav */ 5781286763Smav list_remove(buflist, head); 5782286763Smav list_insert_after(buflist, hdr, head); 5783286763Smav 5784286763Smav mutex_exit(&dev->l2ad_mtx); 5785286763Smav 5786286763Smav /* 5787286763Smav * We wait for the hash lock to become available 5788286763Smav * to try and prevent busy waiting, and increase 5789286763Smav * the chance we'll be able to acquire the lock 5790286763Smav * the next time around. 5791286763Smav */ 5792286763Smav mutex_enter(hash_lock); 5793286763Smav mutex_exit(hash_lock); 5794286763Smav goto top; 5795185029Spjd } 5796185029Spjd 5797286570Smav /* 5798286763Smav * We could not have been moved into the arc_l2c_only 5799286763Smav * state while in-flight due to our ARC_FLAG_L2_WRITING 5800286763Smav * bit being set. Let's just ensure that's being enforced. 5801286570Smav */ 5802286763Smav ASSERT(HDR_HAS_L1HDR(hdr)); 5803286570Smav 5804286763Smav /* 5805286763Smav * We may have allocated a buffer for L2ARC compression, 5806286763Smav * we must release it to avoid leaking this data. 5807286763Smav */ 5808286763Smav l2arc_release_cdata_buf(hdr); 5809286763Smav 5810185029Spjd if (zio->io_error != 0) { 5811185029Spjd /* 5812185029Spjd * Error - drop L2ARC entry. 5813185029Spjd */ 5814286570Smav trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5815286570Smav hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5816286570Smav hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5817286570Smav 5818286570Smav ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5819275811Sdelphij ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5820286598Smav 5821286598Smav bytes_dropped += hdr->b_l2hdr.b_asize; 5822286598Smav (void) refcount_remove_many(&dev->l2ad_alloc, 5823286598Smav hdr->b_l2hdr.b_asize, hdr); 5824185029Spjd } 5825185029Spjd 5826185029Spjd /* 5827286763Smav * Allow ARC to begin reads and ghost list evictions to 5828286763Smav * this L2ARC entry. 5829185029Spjd */ 5830275811Sdelphij hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5831185029Spjd 5832185029Spjd mutex_exit(hash_lock); 5833185029Spjd } 5834185029Spjd 5835185029Spjd atomic_inc_64(&l2arc_writes_done); 5836185029Spjd list_remove(buflist, head); 5837286570Smav ASSERT(!HDR_HAS_L1HDR(head)); 5838286570Smav kmem_cache_free(hdr_l2only_cache, head); 5839286570Smav mutex_exit(&dev->l2ad_mtx); 5840185029Spjd 5841268085Sdelphij vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5842268085Sdelphij 5843185029Spjd l2arc_do_free_on_write(); 5844185029Spjd 5845185029Spjd kmem_free(cb, sizeof (l2arc_write_callback_t)); 5846185029Spjd} 5847185029Spjd 5848185029Spjd/* 5849185029Spjd * A read to a cache device completed. Validate buffer contents before 5850185029Spjd * handing over to the regular ARC routines. 5851185029Spjd */ 5852185029Spjdstatic void 5853185029Spjdl2arc_read_done(zio_t *zio) 5854185029Spjd{ 5855185029Spjd l2arc_read_callback_t *cb; 5856185029Spjd arc_buf_hdr_t *hdr; 5857185029Spjd arc_buf_t *buf; 5858185029Spjd kmutex_t *hash_lock; 5859185029Spjd int equal; 5860185029Spjd 5861185029Spjd ASSERT(zio->io_vd != NULL); 5862185029Spjd ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5863185029Spjd 5864185029Spjd spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5865185029Spjd 5866185029Spjd cb = zio->io_private; 5867185029Spjd ASSERT(cb != NULL); 5868185029Spjd buf = cb->l2rcb_buf; 5869185029Spjd ASSERT(buf != NULL); 5870185029Spjd 5871219089Spjd hash_lock = HDR_LOCK(buf->b_hdr); 5872185029Spjd mutex_enter(hash_lock); 5873219089Spjd hdr = buf->b_hdr; 5874219089Spjd ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5875185029Spjd 5876185029Spjd /* 5877251478Sdelphij * If the buffer was compressed, decompress it first. 5878251478Sdelphij */ 5879251478Sdelphij if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5880251478Sdelphij l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5881251478Sdelphij ASSERT(zio->io_data != NULL); 5882251478Sdelphij 5883251478Sdelphij /* 5884185029Spjd * Check this survived the L2ARC journey. 5885185029Spjd */ 5886185029Spjd equal = arc_cksum_equal(buf); 5887185029Spjd if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5888185029Spjd mutex_exit(hash_lock); 5889185029Spjd zio->io_private = buf; 5890185029Spjd zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5891185029Spjd zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5892185029Spjd arc_read_done(zio); 5893185029Spjd } else { 5894185029Spjd mutex_exit(hash_lock); 5895185029Spjd /* 5896185029Spjd * Buffer didn't survive caching. Increment stats and 5897185029Spjd * reissue to the original storage device. 5898185029Spjd */ 5899185029Spjd if (zio->io_error != 0) { 5900185029Spjd ARCSTAT_BUMP(arcstat_l2_io_error); 5901185029Spjd } else { 5902249195Smm zio->io_error = SET_ERROR(EIO); 5903185029Spjd } 5904185029Spjd if (!equal) 5905185029Spjd ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5906185029Spjd 5907185029Spjd /* 5908185029Spjd * If there's no waiter, issue an async i/o to the primary 5909185029Spjd * storage now. If there *is* a waiter, the caller must 5910185029Spjd * issue the i/o in a context where it's OK to block. 5911185029Spjd */ 5912209962Smm if (zio->io_waiter == NULL) { 5913209962Smm zio_t *pio = zio_unique_parent(zio); 5914209962Smm 5915209962Smm ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 5916209962Smm 5917209962Smm zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 5918185029Spjd buf->b_data, zio->io_size, arc_read_done, buf, 5919185029Spjd zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 5920209962Smm } 5921185029Spjd } 5922185029Spjd 5923185029Spjd kmem_free(cb, sizeof (l2arc_read_callback_t)); 5924185029Spjd} 5925185029Spjd 5926185029Spjd/* 5927185029Spjd * This is the list priority from which the L2ARC will search for pages to 5928185029Spjd * cache. This is used within loops (0..3) to cycle through lists in the 5929185029Spjd * desired order. This order can have a significant effect on cache 5930185029Spjd * performance. 5931185029Spjd * 5932185029Spjd * Currently the metadata lists are hit first, MFU then MRU, followed by 5933185029Spjd * the data lists. This function returns a locked list, and also returns 5934185029Spjd * the lock pointer. 5935185029Spjd */ 5936286763Smavstatic multilist_sublist_t * 5937286763Smavl2arc_sublist_lock(int list_num) 5938185029Spjd{ 5939286763Smav multilist_t *ml = NULL; 5940286763Smav unsigned int idx; 5941185029Spjd 5942286762Smav ASSERT(list_num >= 0 && list_num <= 3); 5943206796Spjd 5944286762Smav switch (list_num) { 5945286762Smav case 0: 5946286763Smav ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 5947286762Smav break; 5948286762Smav case 1: 5949286763Smav ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 5950286762Smav break; 5951286762Smav case 2: 5952286763Smav ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 5953286762Smav break; 5954286762Smav case 3: 5955286763Smav ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 5956286762Smav break; 5957185029Spjd } 5958185029Spjd 5959286763Smav /* 5960286763Smav * Return a randomly-selected sublist. This is acceptable 5961286763Smav * because the caller feeds only a little bit of data for each 5962286763Smav * call (8MB). Subsequent calls will result in different 5963286763Smav * sublists being selected. 5964286763Smav */ 5965286763Smav idx = multilist_get_random_index(ml); 5966286763Smav return (multilist_sublist_lock(ml, idx)); 5967185029Spjd} 5968185029Spjd 5969185029Spjd/* 5970185029Spjd * Evict buffers from the device write hand to the distance specified in 5971185029Spjd * bytes. This distance may span populated buffers, it may span nothing. 5972185029Spjd * This is clearing a region on the L2ARC device ready for writing. 5973185029Spjd * If the 'all' boolean is set, every buffer is evicted. 5974185029Spjd */ 5975185029Spjdstatic void 5976185029Spjdl2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 5977185029Spjd{ 5978185029Spjd list_t *buflist; 5979275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev; 5980185029Spjd kmutex_t *hash_lock; 5981185029Spjd uint64_t taddr; 5982185029Spjd 5983286570Smav buflist = &dev->l2ad_buflist; 5984185029Spjd 5985185029Spjd if (!all && dev->l2ad_first) { 5986185029Spjd /* 5987185029Spjd * This is the first sweep through the device. There is 5988185029Spjd * nothing to evict. 5989185029Spjd */ 5990185029Spjd return; 5991185029Spjd } 5992185029Spjd 5993185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 5994185029Spjd /* 5995185029Spjd * When nearing the end of the device, evict to the end 5996185029Spjd * before the device write hand jumps to the start. 5997185029Spjd */ 5998185029Spjd taddr = dev->l2ad_end; 5999185029Spjd } else { 6000185029Spjd taddr = dev->l2ad_hand + distance; 6001185029Spjd } 6002185029Spjd DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6003185029Spjd uint64_t, taddr, boolean_t, all); 6004185029Spjd 6005185029Spjdtop: 6006286570Smav mutex_enter(&dev->l2ad_mtx); 6007275811Sdelphij for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6008275811Sdelphij hdr_prev = list_prev(buflist, hdr); 6009185029Spjd 6010275811Sdelphij hash_lock = HDR_LOCK(hdr); 6011286763Smav 6012286763Smav /* 6013286763Smav * We cannot use mutex_enter or else we can deadlock 6014286763Smav * with l2arc_write_buffers (due to swapping the order 6015286763Smav * the hash lock and l2ad_mtx are taken). 6016286763Smav */ 6017185029Spjd if (!mutex_tryenter(hash_lock)) { 6018185029Spjd /* 6019185029Spjd * Missed the hash lock. Retry. 6020185029Spjd */ 6021185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6022286570Smav mutex_exit(&dev->l2ad_mtx); 6023185029Spjd mutex_enter(hash_lock); 6024185029Spjd mutex_exit(hash_lock); 6025185029Spjd goto top; 6026185029Spjd } 6027185029Spjd 6028275811Sdelphij if (HDR_L2_WRITE_HEAD(hdr)) { 6029185029Spjd /* 6030185029Spjd * We hit a write head node. Leave it for 6031185029Spjd * l2arc_write_done(). 6032185029Spjd */ 6033275811Sdelphij list_remove(buflist, hdr); 6034185029Spjd mutex_exit(hash_lock); 6035185029Spjd continue; 6036185029Spjd } 6037185029Spjd 6038286570Smav if (!all && HDR_HAS_L2HDR(hdr) && 6039286570Smav (hdr->b_l2hdr.b_daddr > taddr || 6040286570Smav hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6041185029Spjd /* 6042185029Spjd * We've evicted to the target address, 6043185029Spjd * or the end of the device. 6044185029Spjd */ 6045185029Spjd mutex_exit(hash_lock); 6046185029Spjd break; 6047185029Spjd } 6048185029Spjd 6049286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 6050286570Smav if (!HDR_HAS_L1HDR(hdr)) { 6051275811Sdelphij ASSERT(!HDR_L2_READING(hdr)); 6052185029Spjd /* 6053185029Spjd * This doesn't exist in the ARC. Destroy. 6054185029Spjd * arc_hdr_destroy() will call list_remove() 6055185029Spjd * and decrement arcstat_l2_size. 6056185029Spjd */ 6057275811Sdelphij arc_change_state(arc_anon, hdr, hash_lock); 6058275811Sdelphij arc_hdr_destroy(hdr); 6059185029Spjd } else { 6060286570Smav ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6061286570Smav ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6062185029Spjd /* 6063185029Spjd * Invalidate issued or about to be issued 6064185029Spjd * reads, since we may be about to write 6065185029Spjd * over this location. 6066185029Spjd */ 6067275811Sdelphij if (HDR_L2_READING(hdr)) { 6068185029Spjd ARCSTAT_BUMP(arcstat_l2_evict_reading); 6069275811Sdelphij hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6070185029Spjd } 6071185029Spjd 6072286763Smav /* Ensure this header has finished being written */ 6073286763Smav ASSERT(!HDR_L2_WRITING(hdr)); 6074286763Smav ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6075286763Smav 6076286598Smav arc_hdr_l2hdr_destroy(hdr); 6077185029Spjd } 6078185029Spjd mutex_exit(hash_lock); 6079185029Spjd } 6080286570Smav mutex_exit(&dev->l2ad_mtx); 6081185029Spjd} 6082185029Spjd 6083185029Spjd/* 6084185029Spjd * Find and write ARC buffers to the L2ARC device. 6085185029Spjd * 6086275811Sdelphij * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6087185029Spjd * for reading until they have completed writing. 6088251478Sdelphij * The headroom_boost is an in-out parameter used to maintain headroom boost 6089251478Sdelphij * state between calls to this function. 6090251478Sdelphij * 6091251478Sdelphij * Returns the number of bytes actually written (which may be smaller than 6092251478Sdelphij * the delta by which the device hand has changed due to alignment). 6093185029Spjd */ 6094208373Smmstatic uint64_t 6095251478Sdelphijl2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6096251478Sdelphij boolean_t *headroom_boost) 6097185029Spjd{ 6098275811Sdelphij arc_buf_hdr_t *hdr, *hdr_prev, *head; 6099251478Sdelphij uint64_t write_asize, write_psize, write_sz, headroom, 6100251478Sdelphij buf_compress_minsz; 6101185029Spjd void *buf_data; 6102251478Sdelphij boolean_t full; 6103185029Spjd l2arc_write_callback_t *cb; 6104185029Spjd zio_t *pio, *wzio; 6105228103Smm uint64_t guid = spa_load_guid(spa); 6106251478Sdelphij const boolean_t do_headroom_boost = *headroom_boost; 6107185029Spjd int try; 6108185029Spjd 6109185029Spjd ASSERT(dev->l2ad_vdev != NULL); 6110185029Spjd 6111251478Sdelphij /* Lower the flag now, we might want to raise it again later. */ 6112251478Sdelphij *headroom_boost = B_FALSE; 6113251478Sdelphij 6114185029Spjd pio = NULL; 6115251478Sdelphij write_sz = write_asize = write_psize = 0; 6116185029Spjd full = B_FALSE; 6117286570Smav head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6118275811Sdelphij head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6119286570Smav head->b_flags |= ARC_FLAG_HAS_L2HDR; 6120185029Spjd 6121205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6122185029Spjd /* 6123251478Sdelphij * We will want to try to compress buffers that are at least 2x the 6124251478Sdelphij * device sector size. 6125251478Sdelphij */ 6126251478Sdelphij buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6127251478Sdelphij 6128251478Sdelphij /* 6129185029Spjd * Copy buffers for L2ARC writing. 6130185029Spjd */ 6131286762Smav for (try = 0; try <= 3; try++) { 6132286763Smav multilist_sublist_t *mls = l2arc_sublist_lock(try); 6133251478Sdelphij uint64_t passed_sz = 0; 6134251478Sdelphij 6135205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6136185029Spjd 6137185029Spjd /* 6138185029Spjd * L2ARC fast warmup. 6139185029Spjd * 6140185029Spjd * Until the ARC is warm and starts to evict, read from the 6141185029Spjd * head of the ARC lists rather than the tail. 6142185029Spjd */ 6143185029Spjd if (arc_warm == B_FALSE) 6144286763Smav hdr = multilist_sublist_head(mls); 6145185029Spjd else 6146286763Smav hdr = multilist_sublist_tail(mls); 6147275811Sdelphij if (hdr == NULL) 6148205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6149185029Spjd 6150286762Smav headroom = target_sz * l2arc_headroom; 6151251478Sdelphij if (do_headroom_boost) 6152251478Sdelphij headroom = (headroom * l2arc_headroom_boost) / 100; 6153251478Sdelphij 6154275811Sdelphij for (; hdr; hdr = hdr_prev) { 6155251478Sdelphij kmutex_t *hash_lock; 6156251478Sdelphij uint64_t buf_sz; 6157251478Sdelphij 6158185029Spjd if (arc_warm == B_FALSE) 6159286763Smav hdr_prev = multilist_sublist_next(mls, hdr); 6160185029Spjd else 6161286763Smav hdr_prev = multilist_sublist_prev(mls, hdr); 6162275811Sdelphij ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6163206796Spjd 6164275811Sdelphij hash_lock = HDR_LOCK(hdr); 6165251478Sdelphij if (!mutex_tryenter(hash_lock)) { 6166205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6167185029Spjd /* 6168185029Spjd * Skip this buffer rather than waiting. 6169185029Spjd */ 6170185029Spjd continue; 6171185029Spjd } 6172185029Spjd 6173275811Sdelphij passed_sz += hdr->b_size; 6174185029Spjd if (passed_sz > headroom) { 6175185029Spjd /* 6176185029Spjd * Searched too far. 6177185029Spjd */ 6178185029Spjd mutex_exit(hash_lock); 6179205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6180185029Spjd break; 6181185029Spjd } 6182185029Spjd 6183275811Sdelphij if (!l2arc_write_eligible(guid, hdr)) { 6184185029Spjd mutex_exit(hash_lock); 6185185029Spjd continue; 6186185029Spjd } 6187185029Spjd 6188275811Sdelphij if ((write_sz + hdr->b_size) > target_sz) { 6189185029Spjd full = B_TRUE; 6190185029Spjd mutex_exit(hash_lock); 6191205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_full); 6192185029Spjd break; 6193185029Spjd } 6194185029Spjd 6195185029Spjd if (pio == NULL) { 6196185029Spjd /* 6197185029Spjd * Insert a dummy header on the buflist so 6198185029Spjd * l2arc_write_done() can find where the 6199185029Spjd * write buffers begin without searching. 6200185029Spjd */ 6201286763Smav mutex_enter(&dev->l2ad_mtx); 6202286570Smav list_insert_head(&dev->l2ad_buflist, head); 6203286763Smav mutex_exit(&dev->l2ad_mtx); 6204185029Spjd 6205185029Spjd cb = kmem_alloc( 6206185029Spjd sizeof (l2arc_write_callback_t), KM_SLEEP); 6207185029Spjd cb->l2wcb_dev = dev; 6208185029Spjd cb->l2wcb_head = head; 6209185029Spjd pio = zio_root(spa, l2arc_write_done, cb, 6210185029Spjd ZIO_FLAG_CANFAIL); 6211205231Skmacy ARCSTAT_BUMP(arcstat_l2_write_pios); 6212185029Spjd } 6213185029Spjd 6214185029Spjd /* 6215185029Spjd * Create and add a new L2ARC header. 6216185029Spjd */ 6217286570Smav hdr->b_l2hdr.b_dev = dev; 6218275811Sdelphij hdr->b_flags |= ARC_FLAG_L2_WRITING; 6219251478Sdelphij /* 6220251478Sdelphij * Temporarily stash the data buffer in b_tmp_cdata. 6221251478Sdelphij * The subsequent write step will pick it up from 6222286570Smav * there. This is because can't access b_l1hdr.b_buf 6223251478Sdelphij * without holding the hash_lock, which we in turn 6224251478Sdelphij * can't access without holding the ARC list locks 6225251478Sdelphij * (which we want to avoid during compression/writing). 6226251478Sdelphij */ 6227286570Smav HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 6228286570Smav hdr->b_l2hdr.b_asize = hdr->b_size; 6229286570Smav hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6230251478Sdelphij 6231286598Smav /* 6232286598Smav * Explicitly set the b_daddr field to a known 6233286598Smav * value which means "invalid address". This 6234286598Smav * enables us to differentiate which stage of 6235286598Smav * l2arc_write_buffers() the particular header 6236286598Smav * is in (e.g. this loop, or the one below). 6237286598Smav * ARC_FLAG_L2_WRITING is not enough to make 6238286598Smav * this distinction, and we need to know in 6239286598Smav * order to do proper l2arc vdev accounting in 6240286598Smav * arc_release() and arc_hdr_destroy(). 6241286598Smav * 6242286598Smav * Note, we can't use a new flag to distinguish 6243286598Smav * the two stages because we don't hold the 6244286598Smav * header's hash_lock below, in the second stage 6245286598Smav * of this function. Thus, we can't simply 6246286598Smav * change the b_flags field to denote that the 6247286598Smav * IO has been sent. We can change the b_daddr 6248286598Smav * field of the L2 portion, though, since we'll 6249286598Smav * be holding the l2ad_mtx; which is why we're 6250286598Smav * using it to denote the header's state change. 6251286598Smav */ 6252286598Smav hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6253286598Smav 6254275811Sdelphij buf_sz = hdr->b_size; 6255286570Smav hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6256185029Spjd 6257286763Smav mutex_enter(&dev->l2ad_mtx); 6258286570Smav list_insert_head(&dev->l2ad_buflist, hdr); 6259286763Smav mutex_exit(&dev->l2ad_mtx); 6260251478Sdelphij 6261185029Spjd /* 6262185029Spjd * Compute and store the buffer cksum before 6263185029Spjd * writing. On debug the cksum is verified first. 6264185029Spjd */ 6265286570Smav arc_cksum_verify(hdr->b_l1hdr.b_buf); 6266286570Smav arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6267185029Spjd 6268185029Spjd mutex_exit(hash_lock); 6269185029Spjd 6270251478Sdelphij write_sz += buf_sz; 6271251478Sdelphij } 6272251478Sdelphij 6273286763Smav multilist_sublist_unlock(mls); 6274251478Sdelphij 6275251478Sdelphij if (full == B_TRUE) 6276251478Sdelphij break; 6277251478Sdelphij } 6278251478Sdelphij 6279251478Sdelphij /* No buffers selected for writing? */ 6280251478Sdelphij if (pio == NULL) { 6281251478Sdelphij ASSERT0(write_sz); 6282286570Smav ASSERT(!HDR_HAS_L1HDR(head)); 6283286570Smav kmem_cache_free(hdr_l2only_cache, head); 6284251478Sdelphij return (0); 6285251478Sdelphij } 6286251478Sdelphij 6287286763Smav mutex_enter(&dev->l2ad_mtx); 6288286763Smav 6289251478Sdelphij /* 6290251478Sdelphij * Now start writing the buffers. We're starting at the write head 6291251478Sdelphij * and work backwards, retracing the course of the buffer selector 6292251478Sdelphij * loop above. 6293251478Sdelphij */ 6294286570Smav for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6295286570Smav hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6296251478Sdelphij uint64_t buf_sz; 6297251478Sdelphij 6298251478Sdelphij /* 6299286763Smav * We rely on the L1 portion of the header below, so 6300286763Smav * it's invalid for this header to have been evicted out 6301286763Smav * of the ghost cache, prior to being written out. The 6302286763Smav * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6303286763Smav */ 6304286763Smav ASSERT(HDR_HAS_L1HDR(hdr)); 6305286763Smav 6306286763Smav /* 6307251478Sdelphij * We shouldn't need to lock the buffer here, since we flagged 6308275811Sdelphij * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6309275811Sdelphij * take care to only access its L2 cache parameters. In 6310286570Smav * particular, hdr->l1hdr.b_buf may be invalid by now due to 6311275811Sdelphij * ARC eviction. 6312251478Sdelphij */ 6313286570Smav hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6314251478Sdelphij 6315286570Smav if ((HDR_L2COMPRESS(hdr)) && 6316286570Smav hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6317286570Smav if (l2arc_compress_buf(hdr)) { 6318251478Sdelphij /* 6319251478Sdelphij * If compression succeeded, enable headroom 6320251478Sdelphij * boost on the next scan cycle. 6321251478Sdelphij */ 6322251478Sdelphij *headroom_boost = B_TRUE; 6323251478Sdelphij } 6324251478Sdelphij } 6325251478Sdelphij 6326251478Sdelphij /* 6327251478Sdelphij * Pick up the buffer data we had previously stashed away 6328251478Sdelphij * (and now potentially also compressed). 6329251478Sdelphij */ 6330286570Smav buf_data = hdr->b_l1hdr.b_tmp_cdata; 6331286570Smav buf_sz = hdr->b_l2hdr.b_asize; 6332251478Sdelphij 6333274172Savg /* 6334274172Savg * If the data has not been compressed, then clear b_tmp_cdata 6335274172Savg * to make sure that it points only to a temporary compression 6336274172Savg * buffer. 6337274172Savg */ 6338286570Smav if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))) 6339286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 6340274172Savg 6341286598Smav /* 6342286598Smav * We need to do this regardless if buf_sz is zero or 6343286598Smav * not, otherwise, when this l2hdr is evicted we'll 6344286598Smav * remove a reference that was never added. 6345286598Smav */ 6346286598Smav (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6347286598Smav 6348251478Sdelphij /* Compression may have squashed the buffer to zero length. */ 6349251478Sdelphij if (buf_sz != 0) { 6350251478Sdelphij uint64_t buf_p_sz; 6351251478Sdelphij 6352185029Spjd wzio = zio_write_phys(pio, dev->l2ad_vdev, 6353185029Spjd dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6354185029Spjd NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6355185029Spjd ZIO_FLAG_CANFAIL, B_FALSE); 6356185029Spjd 6357185029Spjd DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6358185029Spjd zio_t *, wzio); 6359185029Spjd (void) zio_nowait(wzio); 6360185029Spjd 6361251478Sdelphij write_asize += buf_sz; 6362286598Smav 6363185029Spjd /* 6364185029Spjd * Keep the clock hand suitably device-aligned. 6365185029Spjd */ 6366251478Sdelphij buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6367251478Sdelphij write_psize += buf_p_sz; 6368251478Sdelphij dev->l2ad_hand += buf_p_sz; 6369185029Spjd } 6370251478Sdelphij } 6371185029Spjd 6372286570Smav mutex_exit(&dev->l2ad_mtx); 6373185029Spjd 6374251478Sdelphij ASSERT3U(write_asize, <=, target_sz); 6375185029Spjd ARCSTAT_BUMP(arcstat_l2_writes_sent); 6376251478Sdelphij ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6377185029Spjd ARCSTAT_INCR(arcstat_l2_size, write_sz); 6378251478Sdelphij ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6379275096Sdelphij vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6380185029Spjd 6381185029Spjd /* 6382185029Spjd * Bump device hand to the device start if it is approaching the end. 6383185029Spjd * l2arc_evict() will already have evicted ahead for this case. 6384185029Spjd */ 6385185029Spjd if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6386185029Spjd dev->l2ad_hand = dev->l2ad_start; 6387185029Spjd dev->l2ad_first = B_FALSE; 6388185029Spjd } 6389185029Spjd 6390208373Smm dev->l2ad_writing = B_TRUE; 6391185029Spjd (void) zio_wait(pio); 6392208373Smm dev->l2ad_writing = B_FALSE; 6393208373Smm 6394251478Sdelphij return (write_asize); 6395185029Spjd} 6396185029Spjd 6397185029Spjd/* 6398251478Sdelphij * Compresses an L2ARC buffer. 6399286570Smav * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6400251478Sdelphij * size in l2hdr->b_asize. This routine tries to compress the data and 6401251478Sdelphij * depending on the compression result there are three possible outcomes: 6402251478Sdelphij * *) The buffer was incompressible. The original l2hdr contents were left 6403251478Sdelphij * untouched and are ready for writing to an L2 device. 6404251478Sdelphij * *) The buffer was all-zeros, so there is no need to write it to an L2 6405251478Sdelphij * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6406251478Sdelphij * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6407251478Sdelphij * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6408251478Sdelphij * data buffer which holds the compressed data to be written, and b_asize 6409251478Sdelphij * tells us how much data there is. b_compress is set to the appropriate 6410251478Sdelphij * compression algorithm. Once writing is done, invoke 6411251478Sdelphij * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6412251478Sdelphij * 6413251478Sdelphij * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6414251478Sdelphij * buffer was incompressible). 6415251478Sdelphij */ 6416251478Sdelphijstatic boolean_t 6417286570Smavl2arc_compress_buf(arc_buf_hdr_t *hdr) 6418251478Sdelphij{ 6419251478Sdelphij void *cdata; 6420268075Sdelphij size_t csize, len, rounded; 6421286570Smav ASSERT(HDR_HAS_L2HDR(hdr)); 6422286570Smav l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6423251478Sdelphij 6424286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 6425286570Smav ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 6426286570Smav ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6427251478Sdelphij 6428251478Sdelphij len = l2hdr->b_asize; 6429251478Sdelphij cdata = zio_data_buf_alloc(len); 6430286570Smav ASSERT3P(cdata, !=, NULL); 6431286570Smav csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6432269086Sdelphij cdata, l2hdr->b_asize); 6433251478Sdelphij 6434251478Sdelphij if (csize == 0) { 6435251478Sdelphij /* zero block, indicate that there's nothing to write */ 6436251478Sdelphij zio_data_buf_free(cdata, len); 6437286570Smav HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 6438251478Sdelphij l2hdr->b_asize = 0; 6439286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 6440251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6441251478Sdelphij return (B_TRUE); 6442274628Savg } 6443274628Savg 6444274628Savg rounded = P2ROUNDUP(csize, 6445274628Savg (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6446274628Savg if (rounded < len) { 6447251478Sdelphij /* 6448251478Sdelphij * Compression succeeded, we'll keep the cdata around for 6449251478Sdelphij * writing and release it afterwards. 6450251478Sdelphij */ 6451274628Savg if (rounded > csize) { 6452274628Savg bzero((char *)cdata + csize, rounded - csize); 6453274628Savg csize = rounded; 6454274628Savg } 6455286570Smav HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 6456251478Sdelphij l2hdr->b_asize = csize; 6457286570Smav hdr->b_l1hdr.b_tmp_cdata = cdata; 6458251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_successes); 6459251478Sdelphij return (B_TRUE); 6460251478Sdelphij } else { 6461251478Sdelphij /* 6462251478Sdelphij * Compression failed, release the compressed buffer. 6463251478Sdelphij * l2hdr will be left unmodified. 6464251478Sdelphij */ 6465251478Sdelphij zio_data_buf_free(cdata, len); 6466251478Sdelphij ARCSTAT_BUMP(arcstat_l2_compress_failures); 6467251478Sdelphij return (B_FALSE); 6468251478Sdelphij } 6469251478Sdelphij} 6470251478Sdelphij 6471251478Sdelphij/* 6472251478Sdelphij * Decompresses a zio read back from an l2arc device. On success, the 6473251478Sdelphij * underlying zio's io_data buffer is overwritten by the uncompressed 6474251478Sdelphij * version. On decompression error (corrupt compressed stream), the 6475251478Sdelphij * zio->io_error value is set to signal an I/O error. 6476251478Sdelphij * 6477251478Sdelphij * Please note that the compressed data stream is not checksummed, so 6478251478Sdelphij * if the underlying device is experiencing data corruption, we may feed 6479251478Sdelphij * corrupt data to the decompressor, so the decompressor needs to be 6480251478Sdelphij * able to handle this situation (LZ4 does). 6481251478Sdelphij */ 6482251478Sdelphijstatic void 6483251478Sdelphijl2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6484251478Sdelphij{ 6485251478Sdelphij ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6486251478Sdelphij 6487251478Sdelphij if (zio->io_error != 0) { 6488251478Sdelphij /* 6489251478Sdelphij * An io error has occured, just restore the original io 6490251478Sdelphij * size in preparation for a main pool read. 6491251478Sdelphij */ 6492251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 6493251478Sdelphij return; 6494251478Sdelphij } 6495251478Sdelphij 6496251478Sdelphij if (c == ZIO_COMPRESS_EMPTY) { 6497251478Sdelphij /* 6498251478Sdelphij * An empty buffer results in a null zio, which means we 6499251478Sdelphij * need to fill its io_data after we're done restoring the 6500251478Sdelphij * buffer's contents. 6501251478Sdelphij */ 6502286570Smav ASSERT(hdr->b_l1hdr.b_buf != NULL); 6503286570Smav bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6504286570Smav zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6505251478Sdelphij } else { 6506251478Sdelphij ASSERT(zio->io_data != NULL); 6507251478Sdelphij /* 6508251478Sdelphij * We copy the compressed data from the start of the arc buffer 6509251478Sdelphij * (the zio_read will have pulled in only what we need, the 6510251478Sdelphij * rest is garbage which we will overwrite at decompression) 6511251478Sdelphij * and then decompress back to the ARC data buffer. This way we 6512251478Sdelphij * can minimize copying by simply decompressing back over the 6513251478Sdelphij * original compressed data (rather than decompressing to an 6514251478Sdelphij * aux buffer and then copying back the uncompressed buffer, 6515251478Sdelphij * which is likely to be much larger). 6516251478Sdelphij */ 6517251478Sdelphij uint64_t csize; 6518251478Sdelphij void *cdata; 6519251478Sdelphij 6520251478Sdelphij csize = zio->io_size; 6521251478Sdelphij cdata = zio_data_buf_alloc(csize); 6522251478Sdelphij bcopy(zio->io_data, cdata, csize); 6523251478Sdelphij if (zio_decompress_data(c, cdata, zio->io_data, csize, 6524251478Sdelphij hdr->b_size) != 0) 6525251478Sdelphij zio->io_error = EIO; 6526251478Sdelphij zio_data_buf_free(cdata, csize); 6527251478Sdelphij } 6528251478Sdelphij 6529251478Sdelphij /* Restore the expected uncompressed IO size. */ 6530251478Sdelphij zio->io_orig_size = zio->io_size = hdr->b_size; 6531251478Sdelphij} 6532251478Sdelphij 6533251478Sdelphij/* 6534251478Sdelphij * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6535251478Sdelphij * This buffer serves as a temporary holder of compressed data while 6536251478Sdelphij * the buffer entry is being written to an l2arc device. Once that is 6537251478Sdelphij * done, we can dispose of it. 6538251478Sdelphij */ 6539251478Sdelphijstatic void 6540275811Sdelphijl2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6541251478Sdelphij{ 6542286763Smav enum zio_compress comp = HDR_GET_COMPRESS(hdr); 6543286763Smav 6544286570Smav ASSERT(HDR_HAS_L1HDR(hdr)); 6545286763Smav ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6546286763Smav 6547286763Smav if (comp == ZIO_COMPRESS_OFF) { 6548251478Sdelphij /* 6549286763Smav * In this case, b_tmp_cdata points to the same buffer 6550286763Smav * as the arc_buf_t's b_data field. We don't want to 6551286763Smav * free it, since the arc_buf_t will handle that. 6552286763Smav */ 6553286763Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 6554286763Smav } else if (comp == ZIO_COMPRESS_EMPTY) { 6555286763Smav /* 6556286763Smav * In this case, b_tmp_cdata was compressed to an empty 6557286763Smav * buffer, thus there's nothing to free and b_tmp_cdata 6558286763Smav * should have been set to NULL in l2arc_write_buffers(). 6559286763Smav */ 6560286763Smav ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6561286763Smav } else { 6562286763Smav /* 6563251478Sdelphij * If the data was compressed, then we've allocated a 6564251478Sdelphij * temporary buffer for it, so now we need to release it. 6565251478Sdelphij */ 6566286570Smav ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6567286570Smav zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6568286570Smav hdr->b_size); 6569286570Smav hdr->b_l1hdr.b_tmp_cdata = NULL; 6570251478Sdelphij } 6571251478Sdelphij} 6572251478Sdelphij 6573251478Sdelphij/* 6574185029Spjd * This thread feeds the L2ARC at regular intervals. This is the beating 6575185029Spjd * heart of the L2ARC. 6576185029Spjd */ 6577185029Spjdstatic void 6578185029Spjdl2arc_feed_thread(void *dummy __unused) 6579185029Spjd{ 6580185029Spjd callb_cpr_t cpr; 6581185029Spjd l2arc_dev_t *dev; 6582185029Spjd spa_t *spa; 6583208373Smm uint64_t size, wrote; 6584219089Spjd clock_t begin, next = ddi_get_lbolt(); 6585251478Sdelphij boolean_t headroom_boost = B_FALSE; 6586185029Spjd 6587185029Spjd CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6588185029Spjd 6589185029Spjd mutex_enter(&l2arc_feed_thr_lock); 6590185029Spjd 6591185029Spjd while (l2arc_thread_exit == 0) { 6592185029Spjd CALLB_CPR_SAFE_BEGIN(&cpr); 6593185029Spjd (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6594219089Spjd next - ddi_get_lbolt()); 6595185029Spjd CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6596219089Spjd next = ddi_get_lbolt() + hz; 6597185029Spjd 6598185029Spjd /* 6599185029Spjd * Quick check for L2ARC devices. 6600185029Spjd */ 6601185029Spjd mutex_enter(&l2arc_dev_mtx); 6602185029Spjd if (l2arc_ndev == 0) { 6603185029Spjd mutex_exit(&l2arc_dev_mtx); 6604185029Spjd continue; 6605185029Spjd } 6606185029Spjd mutex_exit(&l2arc_dev_mtx); 6607219089Spjd begin = ddi_get_lbolt(); 6608185029Spjd 6609185029Spjd /* 6610185029Spjd * This selects the next l2arc device to write to, and in 6611185029Spjd * doing so the next spa to feed from: dev->l2ad_spa. This 6612185029Spjd * will return NULL if there are now no l2arc devices or if 6613185029Spjd * they are all faulted. 6614185029Spjd * 6615185029Spjd * If a device is returned, its spa's config lock is also 6616185029Spjd * held to prevent device removal. l2arc_dev_get_next() 6617185029Spjd * will grab and release l2arc_dev_mtx. 6618185029Spjd */ 6619185029Spjd if ((dev = l2arc_dev_get_next()) == NULL) 6620185029Spjd continue; 6621185029Spjd 6622185029Spjd spa = dev->l2ad_spa; 6623185029Spjd ASSERT(spa != NULL); 6624185029Spjd 6625185029Spjd /* 6626219089Spjd * If the pool is read-only then force the feed thread to 6627219089Spjd * sleep a little longer. 6628219089Spjd */ 6629219089Spjd if (!spa_writeable(spa)) { 6630219089Spjd next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6631219089Spjd spa_config_exit(spa, SCL_L2ARC, dev); 6632219089Spjd continue; 6633219089Spjd } 6634219089Spjd 6635219089Spjd /* 6636185029Spjd * Avoid contributing to memory pressure. 6637185029Spjd */ 6638185029Spjd if (arc_reclaim_needed()) { 6639185029Spjd ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6640185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 6641185029Spjd continue; 6642185029Spjd } 6643185029Spjd 6644185029Spjd ARCSTAT_BUMP(arcstat_l2_feeds); 6645185029Spjd 6646251478Sdelphij size = l2arc_write_size(); 6647185029Spjd 6648185029Spjd /* 6649185029Spjd * Evict L2ARC buffers that will be overwritten. 6650185029Spjd */ 6651185029Spjd l2arc_evict(dev, size, B_FALSE); 6652185029Spjd 6653185029Spjd /* 6654185029Spjd * Write ARC buffers. 6655185029Spjd */ 6656251478Sdelphij wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6657208373Smm 6658208373Smm /* 6659208373Smm * Calculate interval between writes. 6660208373Smm */ 6661208373Smm next = l2arc_write_interval(begin, size, wrote); 6662185029Spjd spa_config_exit(spa, SCL_L2ARC, dev); 6663185029Spjd } 6664185029Spjd 6665185029Spjd l2arc_thread_exit = 0; 6666185029Spjd cv_broadcast(&l2arc_feed_thr_cv); 6667185029Spjd CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6668185029Spjd thread_exit(); 6669185029Spjd} 6670185029Spjd 6671185029Spjdboolean_t 6672185029Spjdl2arc_vdev_present(vdev_t *vd) 6673185029Spjd{ 6674185029Spjd l2arc_dev_t *dev; 6675185029Spjd 6676185029Spjd mutex_enter(&l2arc_dev_mtx); 6677185029Spjd for (dev = list_head(l2arc_dev_list); dev != NULL; 6678185029Spjd dev = list_next(l2arc_dev_list, dev)) { 6679185029Spjd if (dev->l2ad_vdev == vd) 6680185029Spjd break; 6681185029Spjd } 6682185029Spjd mutex_exit(&l2arc_dev_mtx); 6683185029Spjd 6684185029Spjd return (dev != NULL); 6685185029Spjd} 6686185029Spjd 6687185029Spjd/* 6688185029Spjd * Add a vdev for use by the L2ARC. By this point the spa has already 6689185029Spjd * validated the vdev and opened it. 6690185029Spjd */ 6691185029Spjdvoid 6692219089Spjdl2arc_add_vdev(spa_t *spa, vdev_t *vd) 6693185029Spjd{ 6694185029Spjd l2arc_dev_t *adddev; 6695185029Spjd 6696185029Spjd ASSERT(!l2arc_vdev_present(vd)); 6697185029Spjd 6698255753Sgibbs vdev_ashift_optimize(vd); 6699255753Sgibbs 6700185029Spjd /* 6701185029Spjd * Create a new l2arc device entry. 6702185029Spjd */ 6703185029Spjd adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6704185029Spjd adddev->l2ad_spa = spa; 6705185029Spjd adddev->l2ad_vdev = vd; 6706219089Spjd adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6707219089Spjd adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6708185029Spjd adddev->l2ad_hand = adddev->l2ad_start; 6709185029Spjd adddev->l2ad_first = B_TRUE; 6710208373Smm adddev->l2ad_writing = B_FALSE; 6711185029Spjd 6712286570Smav mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6713185029Spjd /* 6714185029Spjd * This is a list of all ARC buffers that are still valid on the 6715185029Spjd * device. 6716185029Spjd */ 6717286570Smav list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6718286570Smav offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6719185029Spjd 6720219089Spjd vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6721286598Smav refcount_create(&adddev->l2ad_alloc); 6722185029Spjd 6723185029Spjd /* 6724185029Spjd * Add device to global list 6725185029Spjd */ 6726185029Spjd mutex_enter(&l2arc_dev_mtx); 6727185029Spjd list_insert_head(l2arc_dev_list, adddev); 6728185029Spjd atomic_inc_64(&l2arc_ndev); 6729185029Spjd mutex_exit(&l2arc_dev_mtx); 6730185029Spjd} 6731185029Spjd 6732185029Spjd/* 6733185029Spjd * Remove a vdev from the L2ARC. 6734185029Spjd */ 6735185029Spjdvoid 6736185029Spjdl2arc_remove_vdev(vdev_t *vd) 6737185029Spjd{ 6738185029Spjd l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6739185029Spjd 6740185029Spjd /* 6741185029Spjd * Find the device by vdev 6742185029Spjd */ 6743185029Spjd mutex_enter(&l2arc_dev_mtx); 6744185029Spjd for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6745185029Spjd nextdev = list_next(l2arc_dev_list, dev); 6746185029Spjd if (vd == dev->l2ad_vdev) { 6747185029Spjd remdev = dev; 6748185029Spjd break; 6749185029Spjd } 6750185029Spjd } 6751185029Spjd ASSERT(remdev != NULL); 6752185029Spjd 6753185029Spjd /* 6754185029Spjd * Remove device from global list 6755185029Spjd */ 6756185029Spjd list_remove(l2arc_dev_list, remdev); 6757185029Spjd l2arc_dev_last = NULL; /* may have been invalidated */ 6758185029Spjd atomic_dec_64(&l2arc_ndev); 6759185029Spjd mutex_exit(&l2arc_dev_mtx); 6760185029Spjd 6761185029Spjd /* 6762185029Spjd * Clear all buflists and ARC references. L2ARC device flush. 6763185029Spjd */ 6764185029Spjd l2arc_evict(remdev, 0, B_TRUE); 6765286570Smav list_destroy(&remdev->l2ad_buflist); 6766286570Smav mutex_destroy(&remdev->l2ad_mtx); 6767286598Smav refcount_destroy(&remdev->l2ad_alloc); 6768185029Spjd kmem_free(remdev, sizeof (l2arc_dev_t)); 6769185029Spjd} 6770185029Spjd 6771185029Spjdvoid 6772185029Spjdl2arc_init(void) 6773185029Spjd{ 6774185029Spjd l2arc_thread_exit = 0; 6775185029Spjd l2arc_ndev = 0; 6776185029Spjd l2arc_writes_sent = 0; 6777185029Spjd l2arc_writes_done = 0; 6778185029Spjd 6779185029Spjd mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6780185029Spjd cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6781185029Spjd mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6782185029Spjd mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6783185029Spjd 6784185029Spjd l2arc_dev_list = &L2ARC_dev_list; 6785185029Spjd l2arc_free_on_write = &L2ARC_free_on_write; 6786185029Spjd list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6787185029Spjd offsetof(l2arc_dev_t, l2ad_node)); 6788185029Spjd list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6789185029Spjd offsetof(l2arc_data_free_t, l2df_list_node)); 6790185029Spjd} 6791185029Spjd 6792185029Spjdvoid 6793185029Spjdl2arc_fini(void) 6794185029Spjd{ 6795185029Spjd /* 6796185029Spjd * This is called from dmu_fini(), which is called from spa_fini(); 6797185029Spjd * Because of this, we can assume that all l2arc devices have 6798185029Spjd * already been removed when the pools themselves were removed. 6799185029Spjd */ 6800185029Spjd 6801185029Spjd l2arc_do_free_on_write(); 6802185029Spjd 6803185029Spjd mutex_destroy(&l2arc_feed_thr_lock); 6804185029Spjd cv_destroy(&l2arc_feed_thr_cv); 6805185029Spjd mutex_destroy(&l2arc_dev_mtx); 6806185029Spjd mutex_destroy(&l2arc_free_on_write_mtx); 6807185029Spjd 6808185029Spjd list_destroy(l2arc_dev_list); 6809185029Spjd list_destroy(l2arc_free_on_write); 6810185029Spjd} 6811185029Spjd 6812185029Spjdvoid 6813185029Spjdl2arc_start(void) 6814185029Spjd{ 6815209962Smm if (!(spa_mode_global & FWRITE)) 6816185029Spjd return; 6817185029Spjd 6818185029Spjd (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6819185029Spjd TS_RUN, minclsyspri); 6820185029Spjd} 6821185029Spjd 6822185029Spjdvoid 6823185029Spjdl2arc_stop(void) 6824185029Spjd{ 6825209962Smm if (!(spa_mode_global & FWRITE)) 6826185029Spjd return; 6827185029Spjd 6828185029Spjd mutex_enter(&l2arc_feed_thr_lock); 6829185029Spjd cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6830185029Spjd l2arc_thread_exit = 1; 6831185029Spjd while (l2arc_thread_exit != 0) 6832185029Spjd cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6833185029Spjd mutex_exit(&l2arc_feed_thr_lock); 6834185029Spjd} 6835